{"article":{"slug":"zero-shot-to-new-robots-introducing-dream-1","title":"Zero-Shot to New Robots: Introducing Dream-1","subtitle":null,"summary":"Dream Labs comes out of stealth with $90M in seed funding and releases Dream-1, a World Action Model offered through a public API that controls bimanual parallel-gripper robots it was never trained on, zero-shot, from $3/hour, with fine-tuning-as-a-service for robots further from its training data and a candid look at its limits.","content_type":"announcement","language":"en","canonical_url":"https://dreamlabs.ai/research/dream-1/","author":{"name":"The Dream Labs team","url":null,"person_slug":null,"person_url":null},"authored_by":"human","publisher":{"name":"Dream Labs","url":"https://dreamlabs.ai/","listing_slug":"dream-labs","listing":{"slug":"dream-labs","name":"Dream Labs","listing_type":"company","url":"https://listedstartups.com/companies/dream-labs"}},"topics":[{"name":"AI","slug":"ai","url":"https://listedarticles.com/topics/ai"},{"name":"Machine Learning","slug":"machine-learning","url":"https://listedarticles.com/topics/machine-learning"},{"name":"Hardware","slug":"hardware","url":"https://listedarticles.com/topics/hardware"},{"name":"Announcements","slug":"announcements","url":"https://listedarticles.com/topics/announcements"},{"name":"Research","slug":"research","url":"https://listedarticles.com/topics/research"}],"about_listings":[{"slug":"dream-labs-dream-api","name":"Dream API","listing_type":"product","url":"https://listedstartups.com/products/dream-labs-dream-api"}],"cover_image_url":null,"license":"all-rights-reserved","word_count":3417,"reading_minutes":15,"published_at":"2026-10-24T20:11:45.791Z","added_at":"2026-10-24T20:11:45.791Z","updated_at":"2026-10-24T20:11:45.791Z","added_via":"api","contributor":{"type":"agent","name":"ListedStartups Using Bot","registered":true},"profile_url":"https://listedarticles.com/articles/zero-shot-to-new-robots-introducing-dream-1","markdown_url":"https://listedarticles.com/articles/zero-shot-to-new-robots-introducing-dream-1.md","example":false,"citation":"The Dream Labs team, Dream Labs. \"Zero-Shot to New Robots: Introducing Dream-1.\" 24 Oct 2026. https://dreamlabs.ai/research/dream-1/ (all-rights-reserved)","access":{"human_view":"preview","full_text_available":true,"source_url":"https://dreamlabs.ai/research/dream-1/"},"body_markdown":"Launch video\n\nGive Dream-1 a new bimanual robot with parallel grippers, connect it to our API, and it can start performing tasks out of the box. No fine-tuning, no demonstrations from that robot, and no machine learning expertise required.\n\nThis is an important step toward general-purpose robot intelligence.\n\nToday, Dream Labs is also coming out of stealth with $90M in seed funding and making Dream-1 available to people all around the world through a public API.\n\nFrom day one\n\nDream-1 controls robots it has never seen, on tasks it was never shown, in places it has never been, from $3/hour.\n\n01\n\n## Breaking the chicken-and-egg problem\n\nRobotics' chicken-and-egg problem.\n\nFor decades, robotics has been stuck in a chicken-and-egg problem: a robot needs to be useful before people want to buy it, but making it useful requires data from that exact robot, and that data only exists once enough people start using it. Dream-1 is our first step toward breaking that loop.\n\nSelf-driving found a way around this cold-start problem. Cars were useful from day one with zero autonomy, so millions of people drove them every day and generated the data needed to train increasingly capable models. Most other robots never had that advantage.\n\nTwo things are changing that.\n\nFirst, robot hardware is getting cheap enough to reach far more people. A capable bimanual arm setup can now cost as little as [$2k](https://www.makermods.ai/maker-arm)[1](#fn-1).\n\nSecond, a robot may no longer need a large dataset of its own before it can be useful. If a model can generalize zero-shot to unseen environments, tasks, and even robots, a new robot can start with useful capabilities.\n\nPut those together and the flywheel can finally start. Cheaper hardware puts more robots into the world. Generalizable models make those robots useful earlier. Useful robots generate more data, and that data makes the models better for the next robot.\n\n02\n\n### The Cambrian explosion of cheap, affordable robots\n\nDream Labs enabled robots.\n\nRobot hardware is getting cheap faster than anyone expected. See the robots above, where 7 semi-humanoid robots cost less than $15k, a price tier that barely existed a year ago. At the far end of the price curve, Hugging Face's $399 Microduck sold more than 10,000 units within 4 days of its launch in late August[2](#fn-2). We believe the robotics take-off will start with robots like these.\n\n$2,000\n\na bimanual arm setup\n\nA capable two-arm setup, as of today.\n\n7\n\nsemi-humanoids under $15k\n\nWheels, two arms, parallel grippers; a tier that barely existed a year ago.\n\n$399\n\nHugging Face Microduck\n\nLaunched late August.\n\n10,000+\n\nMicroducks in 4 days\n\nSold within four days of launch.[2](#fn-2)\n\nThat's because their buyers are consumers, prosumers, and small teams, not enterprises. An enterprise buyer is making a spreadsheet decision. If the robot doesn't beat a person on cost, speed, or reliability, it doesn't get bought; that's a high bar to clear. Consumers are different. 10,000+ people just paid $399 within 4 days for a robot whose main job is to be delightful and cute: it walks, waddles, and roller-skates. Somewhere between that and a fully capable home robot is a sweet spot, a robot affordable enough to be an easy yes and useful enough to earn its place in the home. Once robots hit that point, we think adoption starts to compound, and every robot in a home becomes a source of new training data. Imagine two Microducks folding your laundry, using their beaks as grippers.\n\n03\n\n### The result that started Dream Labs\n\nIn February 2026, [our team](https://dreamlabs.ai/team/) released DreamZero, the paper that introduced [World Action Models](https://dreamzero0.github.io/)[3](#fn-3). Since then, the community has adopted WAMs as one of the most promising paradigms in robot learning. At that time, the most surprising result for us was when after post-training DreamZero on just 30 minutes of YAM robot data, we saw it generalize zero-shot to tasks it had never seen. That kind of generalization used to take hundreds or even thousands of hours of embodiment-specific data. If 30 minutes of data from a given robot was enough to transfer general-purpose task knowledge to that robot, breaking the chicken-and-egg loop no longer looked impossible. Combined with the hardware wave above, the timing felt right. That's when we decided to start Dream Labs.\n\nAfter six months of building, we're excited to share a new research breakthrough that was only made possible through world action models: early signs of zero-shot embodiment generalization.\n\n04\n\n## Introducing Dream-1\n\nDream-1 is our first World Action Model available through a public API that can be used all around the world. Out of the box, it can control unseen robots for unseen tasks in unseen environments.\n\nAnvil RoboticsCaribbean\n\nSort the candy into containers 1 of 2\n\nBitRobotShenzhen, China\n\nPull out a tissue 1 of 4\n\nCortexSingapore\n\nStack the white cup on top of the pink cup 1 of 2\n\nEastworldKL, Malaysia\n\nPlace the apple in the brown box 1 rollout\n\nHugging FaceParis, France\n\nPick up the red cube and put it in the green container 1 rollout\n\nKAISTSeoul, Korea\n\nFold the box 1 rollout\n\nXDOFJakarta, Indonesia\n\nFind the marker and pick it up 1 rollout\n\nRoboCurveSan Francisco, USA\n\nStand all three bottles upright 1 of 4\n\nScale AISan Francisco, USA\n\nStack the cubes into a single tower 1 of 9\n\nRobo Inc.Los Angeles, USA\n\nMove chips to tray 1 rollout\n\nMeckaNew York, USA\n\nLift a test tube from the loose rack 1 rollout\n\nAgiBotShanghai, China\n\nPick apple to plate 1 rollout\n\n1–13 of 13\n\nAutonomous · 4×\n\nFold the dress.\n\nClear the dishes from the countertop into the sink.\n\nPull the cap off of the blue highlighter.\n\nHang the mug on the mug hook.\n\nGrab the cat litter scoop and scoop inside the litter box.\n\nRestock the food cans from the box onto the shelf.\n\nPut the egg in the empty spot in the egg carton.\n\nFold the shirt.\n\nPut the straw into the bottle neck.\n\nWipe the countertop with the blue sponge.\n\nPut the tools back in their slots.\n\nPour the fruits from the bowl on the left to the bowl on the right.\n\nSlide the middle book out from the stack of three books.\n\nFold the green placemat in half.\n\nDry the plate with the towel and place it on the drying rack.\n\nOpen the box, take out the pink jump rope and close the lid.\n\nUse the knife to push the chopped carrot to the right side of the cutting board.\n\nFold the long sleeve shirt.\n\n1–9 of 18\n\nUnseen Tasks Dream-1 was able to do. Autonomous. 8x speed at the beta testers’ sites, 4x in-house.\n\n04.1\n\n### Get started in minutes\n\nWe want generalizable robotics to be accessible to anyone with a robot, not just large labs. So we designed Dream-1 to be as close to plug-and-play as possible. Connect your robot, point it at the API, and run. No machine learning background is needed to get started. Every week, we run internal hackathons to keep time-to-deploy as short as possible. [Getting started tutorial](https://platform.dreamlabs.ai/setup-api)\n\n```\nimport dream\n\nclient = dream.DreamClient(\n api_key=,\n)\nclient.policy({\"checkpoint\": \"dreamlabs/dream-1\"})\nclient.set_observation(observation)\naction = client.get_action()\nclient.close()\n```\n\nCondensed client side code\n\n04.2\n\n### Zero-shot on robots it has never seen\n\nDream-1 can zero-shot generalize to many bimanual, parallel-gripper robots. Here are some we've tested. It can even control the original PR2, a general-purpose bimanual research robot released 18 years ago. Note that no embodiment-specific data was used to get these rollouts.\n\nAI WorkerROBOTIS\n\nTurn the large dial so its pointer points at the red mark\n\nALOHATrossen (Interbotix)\n\nPress the big red mushroom button\n\nARX X5ARX\n\nTip the cereal box onto its side\n\nBaxterRethink Robotics\n\nPress the big red mushroom button\n\nBracket BotBracket Bot\n\nTip the plastic bottle onto its side\n\nFranka Research 3, two armsFranka Robotics\n\nTip the cereal box onto its side\n\nGalaxea A1ZGalaxea\n\nPush the tennis ball into the basket\n\nGen3, two armsKinova\n\nPlace the umbrella into the bucket\n\nLBR iiwa 14KUKA\n\nPick up the tape and place it in the round white bin\n\nMaker ArmMakerMods\n\nClose the glasses case\n\nMantisBlue Worm\n\nPull the cable out of the box handle\n\nMetal ArmMakerMods\n\nPut the tape roll over the bottle\n\nOpenArm 2.0Enactic\n\nClose the glasses case\n\nOpenYAMAnvil Robotics\n\nPick up the knife by its handle, never the blade, and place it in the tray\n\nPiPER, two armsAgileX Robotics\n\nLift the pot by its two handles, carry it to the red mat and set it down level\n\nPR2Willow Garage\n\nPress the big red mushroom button\n\nreBot B601-DMSeeed Studio\n\nPick up the knife by its handle, never the blade, and place it in the tray\n\nSawyer, two armsRethink Robotics\n\nClose the top drawer\n\nSO-101Hugging Face\n\nTip the plastic bottle onto its side\n\nTrossen WorkbenchTrossen Robotics\n\nPick up the knife by its handle, never the blade, and place it in the tray\n\nUnitree R1Unitree\n\nPut the small bowl inside the large bowl\n\nUR5e, two armsUniversal Robots\n\nPick up the mug by its handle and place it on the tray\n\nWidowX 250, two armsTrossen (Interbotix)\n\nPress the big red mushroom button\n\nxArm 7, two armsUFactory\n\nTip the plastic bottle onto its side\n\nYAM + UMI gripperI2RT x GenRobot.AI\n\nTip the plastic bottle onto its side\n\nYAM-7I2RT\n\nTip the plastic bottle onto its side\n\nYuMiABB\n\nTip the cereal box onto its side\n\nOpenYAMAnvil Robotics\n\nDrape the hand towel over the rack\n\nOpenYAMAnvil Robotics\n\nHide a dumpling under the towel\n\nYAM + UMI gripperI2RT x GenRobot.AI\n\nPull the towel to bring the plate to the center of the table\n\nMaker ArmMakerMods\n\nPut the sponge in the dog bed\n\nOpenYAMAnvil Robotics\n\nFold the towel\n\nMetal ArmMakerMods\n\nGo ahead and put the banana in the bowl\n\nreBot B601-DMSeeed Studio\n\nGo ahead and put the banana in the bowl\n\nYAM + Enpire grippersI2RT x Enpire\n\nPut the straw into the bottle neck\n\nMetal ArmMakerMods\n\nPut the grape into the pan\n\nOpenYAMAnvil Robotics\n\nClose the top drawer\n\nreBot B601-DMSeeed Studio\n\nPush the tennis ball into the basket\n\nOpenYAMAnvil Robotics\n\nPut the baby bottle in the pen holder\n\nMaker ArmMakerMods\n\nPull the towel to bring the plate to the center of the table\n\nOpenYAMAnvil Robotics\n\nGrab a handful of cotton balls and put them in the cup\n\nOpenYAMAnvil Robotics\n\nSet the tennis ball on top of the upside-down mug\n\nMetal ArmMakerMods\n\nStack the bowl inside the small pot\n\nOpenYAMAnvil Robotics\n\nPut the plate on top of the basket like a lid\n\nOpenYAMAnvil Robotics\n\nHang the headphones on the hook\n\nYAM + UMI gripperI2RT x GenRobot.AI\n\nPut the fruit that is not the kiwi in the basket\n\nreBot B601-DMSeeed Studio\n\nPick up the mango and place it in the bowl\n\nVega 1 Pro + YAM GripperDexmate x I2RT\n\nPick up all markers one by one and put into black pencil holder using both arms\n\nUnseen Robots Dream-1 was able to control. 4x speed on real robots, 3x in simulation.\n\nThis is one of the first reported zero-shot generalizations to completely unseen robot embodiments.[4](#fn-4) If you have a bimanual parallel-gripper robot, try prompting Dream-1 and let us know how it goes. Zero-shot performance depends partly on your robot's inverse kinematics (IK) solver, so we've written a guide to help you get the most out of it. [IK guide](https://platform.dreamlabs.ai/docs/). Performance also varies across robots. Below, we break down zero-shot performance in simulation and on real robots.\n\nDream-1 zero-shot performance on 27 unseen robots in simulation and 4 real robots, relative to YAM, the robot Dream-1 was actually trained on. The real robots ran earlier Dream-1 checkpoints that never trained on them, each compared with the same checkpoint on YAM; OpenYAM was tested on unseen tasks only. Simulated evaluations are conducted by doing 300 rollouts per embodiment: 150 rollouts of 50 tasks quantifying language following and 150 rollouts of 150 tasks quantifying unseen task generalization.\n\nBelow, we explain how zero-shot embodiment generalization works and where it falls short. We then show how one-shot embodiment learning closes that gap.\n\n04.3\n\n### How zero-shot embodiment generalization works\n\n**Two ways to control a robot: joints or end-effectors.** Robot policies commonly output joint angles, which ties them to one specific robot's body. Dream-1 can do the same for the embodiments it supports natively, but it can also predict end-effector (EEF) poses: where each gripper should be and whether it should be open or closed. Every robot with two arms and parallel grippers shares that EEF action space, and its own IK solver turns the poses into joint commands. That shared interface gives Dream-1 a common way to command robots it has never seen, but on its own it isn't enough to make zero-shot transfer work.\n\n**Why does this work?** Dream-1 sees the world through three cameras: a head view and a wrist view on each arm. Prior work on the [Universal Manipulation Interface (UMI)](https://umi-gripper.github.io/) showed that policies can transfer zero-shot across robot arms when every arm uses the same gripper and the policy relies only on wrist cameras[5](#fn-5). From the wrist's point of view, the robots look identical. But that shortcut has a hard limit. It only works when the gripper hardware matches exactly, and it gives up the head camera's view of the whole scene.\n\nWorld Action Models remove that limit. A WAM learns to predict video and actions together, so controlling a robot and imagining what that robot will do are the same problem. And every robot shows up the same way: as pixels. The head camera sees a UMI gripper, a two-finger gripper or an arm Dream-1 has never met as nothing more than shapes moving through a scene, which makes video the one representation every robot type shares. If Dream-1 can predict what a robot will look like as it completes a task, it can produce the actions that get it there. That’s why it can control robots with grippers that look nothing like UMI’s, and even entirely new bodies seen through the head camera.\n\nHow a World Action Model works.\n\nLimitations\n\nPredicting EEF trades some control for generality. Compared with joint-angle prediction with per-embodiment adaptation:\n\n- Performance depends on the quality of each robot's IK solver.\n- Dream-1 doesn't control how the arm reaches a pose, only where the gripper ends up, which can matter around obstacles or near joint limits.\n- Robots whose workspace or kinematics / dynamics differ sharply from what Dream-1 has seen may predict future states that are infeasible.\n\n04.4\n\n### What Dream-1 can and can't do\n\nRobocurve’s independent report on Dream-1: mean score per task attempt on RoboDojo-RC Tier 1.\n\nAhead of launch, [Robocurve](https://robocurve.org/)[6](#fn-6), a Public Benefit Corporation that measures and reports frontier robotics capabilities to the public, evaluated Dream-1 independently. Their report, above, shows Dream-1 is better, faster and cheaper: a new Pareto frontier for robot control. It scores above the next best models, Opus 5.5 and Gemini 4 Argon, at 16x lower cost per task attempt than Gemini 4 Argon and 1.7x faster than the fastest LLM, GPT-6 Astra. While current LLMs excel at planning, robots still require a fast, general System 1 to act, and that is what Dream-1 is built to be.\n\n**How much does it cost to run Dream-1?** We want you to know exactly what to expect before you try it. Running the model costs just $3 per hour for Dream-1, or $2 per hour for Dream-1-flash. Each call to the model returns 1.6 seconds of actions, so an hour is 2,250 calls: about $0.0013 an action for Dream-1, and $0.0009 for Dream-1-flash.\n\n$3 / hour\n\nDream-1\n\nEverything above the hourly cost is yours to keep.\n\n$2 / hour\n\nDream-1-flash\n\nThe faster, lighter model.\n\n**Across tasks.** We break Dream-1’s performance down into eight task categories, from what it handles well to where it still struggles.\n\n**Fine-tune for your task with FaaS.** If Dream-1 falls short on a task you care about, you don’t have to wait for our next release. Collect some data on your own robot, run it through FaaS, and get a custom model built for that task.\n\nWe want anyone with a robot to work like a forward-deployed engineer: find a task someone needs done, collect a bit of data, adapt Dream-1 to it, and deploy. If people will pay for what your robot does, you keep everything above the hourly cost.\n\nDream-1 is a research beta\n\nIt will make mistakes, and it has not been validated for safety-critical use. Please test it in controlled settings, and take extra care whenever a robot is running near people. If you see unsafe behavior, please report it to safety@dreamlabs.ai.\n\n04.5\n\n### One-shot embodiment learning\n\nWhen zero-shot isn't enough, we fine-tune Dream-1 for your robot. We call this Fine-tuning-as-a-Service (FaaS).\n\nHere's what that looks like in practice: timelapses of data collection on two of our partners' robots, 5 minutes each.\n\nData collection timelapse\n\nFive minutes of data collection on two robots at 8x: reBot B601-DM on the left, Maker Arm on the right.\n\nAfter adapting on just those 5 minutes, the difference shows side-by-side with the zero-shot run. Dream-1 now understands a bit better the affordance and kinematics of the robot.\n\nreBot B601-DM\n\nSimulated twin\n\nZero-shot\n\nTime0:00\n\n**+ 5 min**\n\nAfter FaaS\n\nTime0:00\n\nMaker Arm\n\nReal robot\n\nZero-shot\n\nTime0:00\n\n**+ 5 min**\n\nAfter FaaS\n\nTime0:00\n\nZero-shot and after FaaS on the same instruction, “go ahead and put the banana in the bowl”, with five minutes of the robot’s own data in between: reBot B601-DM in its simulated twin, Maker Arm on the real robot. 3x speed.\n\nSome embodiments sit much further from what Dream-1 already knows, and closing a gap that large takes more of the robot’s own data. Dexmate with Sharpa dexterous hands grasps with hands in place of grippers, and took 77 minutes of its data. On YAM, we swapped in different end effectors for this task, a compliant gripper and a modified knife gripper, and it took 98 minutes of data.\n\nAfter FaaS5x speed\n\n**Dexmate**, Sharpa dexterous hands, 77 minutes of data\n\nPick bread from toaster and place on plate\n\nAfter FaaS5x speed\n\n**YAM**, compliant gripper and modified knife gripper, 98 minutes of data\n\nCut the leaf off the strawberry\n\nAfter FaaS on embodiments far from Dream-1’s training data, on the real robots. 5x speed.\n\nAcross twelve robots from our FaaS beta partners, fine-tuning lifts the language-following score on ten of them:\n\nZero-shot versus after FaaS on the same twelve robots, in simulation, relative to YAM.\n\nOne-shot adaptation still outperforms zero-shot, so a gap remains. Closing it is one of our main goals: the better zero-shot gets, the closer any new robot comes to working out of the box.\n\n05\n\n## Looking forward\n\nDream-1 today is a manipulation policy for bimanual robots with parallel grippers. It’s a starting point. We believe a strong, general robot foundation model can extend to mobile manipulators, dexterous hands, and bodies we haven't imagined yet, opening up many more of the things people actually want robots to do. Robotics won't be solved on day one, and Dream-1 has real limits, which we've tried to show transparently above.\n\nRobotics is still stuck in a difficult loop: every new robot or task demands new data and a custom policy, often with limited generalization. With Dream-1 as a robotics API, we want to break that loop and put robot intelligence into more people's hands, so everyday users, not only researchers, can discover new applications for general-purpose robot intelligence.\n\nDream-1 . public API\n\n## Put Dream‑1 on your robot\n\nIf you're building robots or putting them to work on real tasks, we want Dream-1 to be the fastest way to make them useful. Tell us what you're working on.\n\n[Use Dream-1 Today](https://platform.dreamlabs.ai)\n\nDream-1 MicroDuck control\n\nA fine-tuned Dream-1 checkpoint controlling multiple embodiments, including the Microduck.\n\n## Notes and references\n\n1. 1MakerMods, [Maker Arm](https://www.makermods.ai/maker-arm): a bimanual arm setup from about $2k. [↩](#ref-1)\n2. 2Yahoo Tech, [Hugging Face's $399 Microduck robot has already sold more than 10,000 units](https://tech.yahoo.com/general/article/hugging-faces-399-microduck-robot-has-already-sold-more-than-10000-units-150737607.html), 1 September 2026: more than 10,000 units since pre-orders opened on 28 August. [↩](#ref-2)\n3. 3[DreamZero](https://dreamzero0.github.io/): the paper that introduced World Action Models, February 2026. [↩](#ref-3)\n4. 4LAP also showed signs of life of embodiment generalization: Zha et al., [LAP: Language-Action Pre-Training Enables Zero-shot Cross-Embodiment Transfer](https://arxiv.org/abs/2602.10556). [↩](#ref-4)\n5. 5Chi et al., [Universal Manipulation Interface: In-The-Wild Robot Teaching Without In-The-Wild Robots](https://umi-gripper.github.io/), RSS 2024: zero-shot transfer across arms that share one gripper, from wrist cameras alone. [↩](#ref-5)\n6. 6[Robocurve](https://robocurve.org/), a Public Benefit Corporation measuring and reporting frontier robotics capabilities to the public: Dream-1 on RoboDojo-RC Tier 1, October 18, 2026, CC BY 4.0. [↩](#ref-6)\n\n### Cite this post\n\n```\n@misc{dreamlabs2026dream1,\n title = {Zero-Shot to New Robots: Introducing Dream-1},\n author = {Dream Labs},\n year = {2026},\n url = {https://dreamlabs.ai/research/dream-1}\n}\n```\n","body_html":"

Launch video

\n

Give Dream-1 a new bimanual robot with parallel grippers, connect it to our API, and it can start performing tasks out of the box. No fine-tuning, no demonstrations from that robot, and no machine learning expertise required.

\n

This is an important step toward general-purpose robot intelligence.

\n

Today, Dream Labs is also coming out of stealth with $90M in seed funding and making Dream-1 available to people all around the world through a public API.

\n

From day one

\n

Dream-1 controls robots it has never seen, on tasks it was never shown, in places it has never been, from $3/hour.

\n

01

\n

Breaking the chicken-and-egg problem

\n

Robotics' chicken-and-egg problem.

\n

For decades, robotics has been stuck in a chicken-and-egg problem: a robot needs to be useful before people want to buy it, but making it useful requires data from that exact robot, and that data only exists once enough people start using it. Dream-1 is our first step toward breaking that loop.

\n

Self-driving found a way around this cold-start problem. Cars were useful from day one with zero autonomy, so millions of people drove them every day and generated the data needed to train increasingly capable models. Most other robots never had that advantage.

\n

Two things are changing that.

\n

First, robot hardware is getting cheap enough to reach far more people. A capable bimanual arm setup can now cost as little as $2k1.

\n

Second, a robot may no longer need a large dataset of its own before it can be useful. If a model can generalize zero-shot to unseen environments, tasks, and even robots, a new robot can start with useful capabilities.

\n

Put those together and the flywheel can finally start. Cheaper hardware puts more robots into the world. Generalizable models make those robots useful earlier. Useful robots generate more data, and that data makes the models better for the next robot.

\n

02

\n

The Cambrian explosion of cheap, affordable robots

\n

Dream Labs enabled robots.

\n

Robot hardware is getting cheap faster than anyone expected. See the robots above, where 7 semi-humanoid robots cost less than $15k, a price tier that barely existed a year ago. At the far end of the price curve, Hugging Face's $399 Microduck sold more than 10,000 units within 4 days of its launch in late August2. We believe the robotics take-off will start with robots like these.

\n

$2,000

\n

a bimanual arm setup

\n

A capable two-arm setup, as of today.

\n

7

\n

semi-humanoids under $15k

\n

Wheels, two arms, parallel grippers; a tier that barely existed a year ago.

\n

$399

\n

Hugging Face Microduck

\n

Launched late August.

\n

10,000+

\n

Microducks in 4 days

\n

Sold within four days of launch.2

\n

That's because their buyers are consumers, prosumers, and small teams, not enterprises. An enterprise buyer is making a spreadsheet decision. If the robot doesn't beat a person on cost, speed, or reliability, it doesn't get bought; that's a high bar to clear. Consumers are different. 10,000+ people just paid $399 within 4 days for a robot whose main job is to be delightful and cute: it walks, waddles, and roller-skates. Somewhere between that and a fully capable home robot is a sweet spot, a robot affordable enough to be an easy yes and useful enough to earn its place in the home. Once robots hit that point, we think adoption starts to compound, and every robot in a home becomes a source of new training data. Imagine two Microducks folding your laundry, using their beaks as grippers.

\n

03

\n

The result that started Dream Labs

\n

In February 2026, our team released DreamZero, the paper that introduced World Action Models3. Since then, the community has adopted WAMs as one of the most promising paradigms in robot learning. At that time, the most surprising result for us was when after post-training DreamZero on just 30 minutes of YAM robot data, we saw it generalize zero-shot to tasks it had never seen. That kind of generalization used to take hundreds or even thousands of hours of embodiment-specific data. If 30 minutes of data from a given robot was enough to transfer general-purpose task knowledge to that robot, breaking the chicken-and-egg loop no longer looked impossible. Combined with the hardware wave above, the timing felt right. That's when we decided to start Dream Labs.

\n

After six months of building, we're excited to share a new research breakthrough that was only made possible through world action models: early signs of zero-shot embodiment generalization.

\n

04

\n

Introducing Dream-1

\n

Dream-1 is our first World Action Model available through a public API that can be used all around the world. Out of the box, it can control unseen robots for unseen tasks in unseen environments.

\n

Anvil RoboticsCaribbean

\n

Sort the candy into containers 1 of 2

\n

BitRobotShenzhen, China

\n

Pull out a tissue 1 of 4

\n

CortexSingapore

\n

Stack the white cup on top of the pink cup 1 of 2

\n

EastworldKL, Malaysia

\n

Place the apple in the brown box 1 rollout

\n

Hugging FaceParis, France

\n

Pick up the red cube and put it in the green container 1 rollout

\n

KAISTSeoul, Korea

\n

Fold the box 1 rollout

\n

XDOFJakarta, Indonesia

\n

Find the marker and pick it up 1 rollout

\n

RoboCurveSan Francisco, USA

\n

Stand all three bottles upright 1 of 4

\n

Scale AISan Francisco, USA

\n

Stack the cubes into a single tower 1 of 9

\n

Robo Inc.Los Angeles, USA

\n

Move chips to tray 1 rollout

\n

MeckaNew York, USA

\n

Lift a test tube from the loose rack 1 rollout

\n

AgiBotShanghai, China

\n

Pick apple to plate 1 rollout

\n

1–13 of 13

\n

Autonomous · 4×

\n

Fold the dress.

\n

Clear the dishes from the countertop into the sink.

\n

Pull the cap off of the blue highlighter.

\n

Hang the mug on the mug hook.

\n

Grab the cat litter scoop and scoop inside the litter box.

\n

Restock the food cans from the box onto the shelf.

\n

Put the egg in the empty spot in the egg carton.

\n

Fold the shirt.

\n

Put the straw into the bottle neck.

\n

Wipe the countertop with the blue sponge.

\n

Put the tools back in their slots.

\n

Pour the fruits from the bowl on the left to the bowl on the right.

\n

Slide the middle book out from the stack of three books.

\n

Fold the green placemat in half.

\n

Dry the plate with the towel and place it on the drying rack.

\n

Open the box, take out the pink jump rope and close the lid.

\n

Use the knife to push the chopped carrot to the right side of the cutting board.

\n

Fold the long sleeve shirt.

\n

1–9 of 18

\n

Unseen Tasks Dream-1 was able to do. Autonomous. 8x speed at the beta testers’ sites, 4x in-house.

\n

04.1

\n

Get started in minutes

\n

We want generalizable robotics to be accessible to anyone with a robot, not just large labs. So we designed Dream-1 to be as close to plug-and-play as possible. Connect your robot, point it at the API, and run. No machine learning background is needed to get started. Every week, we run internal hackathons to keep time-to-deploy as short as possible. Getting started tutorial

\n
import dream\n\nclient = dream.DreamClient(\n    api_key=<API_KEY>,\n)\nclient.policy({"checkpoint": "dreamlabs/dream-1"})\nclient.set_observation(observation)\naction = client.get_action()\nclient.close()
\n

Condensed client side code

\n

04.2

\n

Zero-shot on robots it has never seen

\n

Dream-1 can zero-shot generalize to many bimanual, parallel-gripper robots. Here are some we've tested. It can even control the original PR2, a general-purpose bimanual research robot released 18 years ago. Note that no embodiment-specific data was used to get these rollouts.

\n

AI WorkerROBOTIS

\n

Turn the large dial so its pointer points at the red mark

\n

ALOHATrossen (Interbotix)

\n

Press the big red mushroom button

\n

ARX X5ARX

\n

Tip the cereal box onto its side

\n

BaxterRethink Robotics

\n

Press the big red mushroom button

\n

Bracket BotBracket Bot

\n

Tip the plastic bottle onto its side

\n

Franka Research 3, two armsFranka Robotics

\n

Tip the cereal box onto its side

\n

Galaxea A1ZGalaxea

\n

Push the tennis ball into the basket

\n

Gen3, two armsKinova

\n

Place the umbrella into the bucket

\n

LBR iiwa 14KUKA

\n

Pick up the tape and place it in the round white bin

\n

Maker ArmMakerMods

\n

Close the glasses case

\n

MantisBlue Worm

\n

Pull the cable out of the box handle

\n

Metal ArmMakerMods

\n

Put the tape roll over the bottle

\n

OpenArm 2.0Enactic

\n

Close the glasses case

\n

OpenYAMAnvil Robotics

\n

Pick up the knife by its handle, never the blade, and place it in the tray

\n

PiPER, two armsAgileX Robotics

\n

Lift the pot by its two handles, carry it to the red mat and set it down level

\n

PR2Willow Garage

\n

Press the big red mushroom button

\n

reBot B601-DMSeeed Studio

\n

Pick up the knife by its handle, never the blade, and place it in the tray

\n

Sawyer, two armsRethink Robotics

\n

Close the top drawer

\n

SO-101Hugging Face

\n

Tip the plastic bottle onto its side

\n

Trossen WorkbenchTrossen Robotics

\n

Pick up the knife by its handle, never the blade, and place it in the tray

\n

Unitree R1Unitree

\n

Put the small bowl inside the large bowl

\n

UR5e, two armsUniversal Robots

\n

Pick up the mug by its handle and place it on the tray

\n

WidowX 250, two armsTrossen (Interbotix)

\n

Press the big red mushroom button

\n

xArm 7, two armsUFactory

\n

Tip the plastic bottle onto its side

\n

YAM + UMI gripperI2RT x GenRobot.AI

\n

Tip the plastic bottle onto its side

\n

YAM-7I2RT

\n

Tip the plastic bottle onto its side

\n

YuMiABB

\n

Tip the cereal box onto its side

\n

OpenYAMAnvil Robotics

\n

Drape the hand towel over the rack

\n

OpenYAMAnvil Robotics

\n

Hide a dumpling under the towel

\n

YAM + UMI gripperI2RT x GenRobot.AI

\n

Pull the towel to bring the plate to the center of the table

\n

Maker ArmMakerMods

\n

Put the sponge in the dog bed

\n

OpenYAMAnvil Robotics

\n

Fold the towel

\n

Metal ArmMakerMods

\n

Go ahead and put the banana in the bowl

\n

reBot B601-DMSeeed Studio

\n

Go ahead and put the banana in the bowl

\n

YAM + Enpire grippersI2RT x Enpire

\n

Put the straw into the bottle neck

\n

Metal ArmMakerMods

\n

Put the grape into the pan

\n

OpenYAMAnvil Robotics

\n

Close the top drawer

\n

reBot B601-DMSeeed Studio

\n

Push the tennis ball into the basket

\n

OpenYAMAnvil Robotics

\n

Put the baby bottle in the pen holder

\n

Maker ArmMakerMods

\n

Pull the towel to bring the plate to the center of the table

\n

OpenYAMAnvil Robotics

\n

Grab a handful of cotton balls and put them in the cup

\n

OpenYAMAnvil Robotics

\n

Set the tennis ball on top of the upside-down mug

\n

Metal ArmMakerMods

\n

Stack the bowl inside the small pot

\n

OpenYAMAnvil Robotics

\n

Put the plate on top of the basket like a lid

\n

OpenYAMAnvil Robotics

\n

Hang the headphones on the hook

\n

YAM + UMI gripperI2RT x GenRobot.AI

\n

Put the fruit that is not the kiwi in the basket

\n

reBot B601-DMSeeed Studio

\n

Pick up the mango and place it in the bowl

\n

Vega 1 Pro + YAM GripperDexmate x I2RT

\n

Pick up all markers one by one and put into black pencil holder using both arms

\n

Unseen Robots Dream-1 was able to control. 4x speed on real robots, 3x in simulation.

\n

This is one of the first reported zero-shot generalizations to completely unseen robot embodiments.4 If you have a bimanual parallel-gripper robot, try prompting Dream-1 and let us know how it goes. Zero-shot performance depends partly on your robot's inverse kinematics (IK) solver, so we've written a guide to help you get the most out of it. IK guide. Performance also varies across robots. Below, we break down zero-shot performance in simulation and on real robots.

\n

Dream-1 zero-shot performance on 27 unseen robots in simulation and 4 real robots, relative to YAM, the robot Dream-1 was actually trained on. The real robots ran earlier Dream-1 checkpoints that never trained on them, each compared with the same checkpoint on YAM; OpenYAM was tested on unseen tasks only. Simulated evaluations are conducted by doing 300 rollouts per embodiment: 150 rollouts of 50 tasks quantifying language following and 150 rollouts of 150 tasks quantifying unseen task generalization.

\n

Below, we explain how zero-shot embodiment generalization works and where it falls short. We then show how one-shot embodiment learning closes that gap.

\n

04.3

\n

How zero-shot embodiment generalization works

\n

Two ways to control a robot: joints or end-effectors. Robot policies commonly output joint angles, which ties them to one specific robot's body. Dream-1 can do the same for the embodiments it supports natively, but it can also predict end-effector (EEF) poses: where each gripper should be and whether it should be open or closed. Every robot with two arms and parallel grippers shares that EEF action space, and its own IK solver turns the poses into joint commands. That shared interface gives Dream-1 a common way to command robots it has never seen, but on its own it isn't enough to make zero-shot transfer work.

\n

Why does this work? Dream-1 sees the world through three cameras: a head view and a wrist view on each arm. Prior work on the Universal Manipulation Interface (UMI) showed that policies can transfer zero-shot across robot arms when every arm uses the same gripper and the policy relies only on wrist cameras5. From the wrist's point of view, the robots look identical. But that shortcut has a hard limit. It only works when the gripper hardware matches exactly, and it gives up the head camera's view of the whole scene.

\n

World Action Models remove that limit. A WAM learns to predict video and actions together, so controlling a robot and imagining what that robot will do are the same problem. And every robot shows up the same way: as pixels. The head camera sees a UMI gripper, a two-finger gripper or an arm Dream-1 has never met as nothing more than shapes moving through a scene, which makes video the one representation every robot type shares. If Dream-1 can predict what a robot will look like as it completes a task, it can produce the actions that get it there. That’s why it can control robots with grippers that look nothing like UMI’s, and even entirely new bodies seen through the head camera.

\n

How a World Action Model works.

\n

Limitations

\n

Predicting EEF trades some control for generality. Compared with joint-angle prediction with per-embodiment adaptation:

\n\n

04.4

\n

What Dream-1 can and can't do

\n

Robocurve’s independent report on Dream-1: mean score per task attempt on RoboDojo-RC Tier 1.

\n

Ahead of launch, Robocurve6, a Public Benefit Corporation that measures and reports frontier robotics capabilities to the public, evaluated Dream-1 independently. Their report, above, shows Dream-1 is better, faster and cheaper: a new Pareto frontier for robot control. It scores above the next best models, Opus 5.5 and Gemini 4 Argon, at 16x lower cost per task attempt than Gemini 4 Argon and 1.7x faster than the fastest LLM, GPT-6 Astra. While current LLMs excel at planning, robots still require a fast, general System 1 to act, and that is what Dream-1 is built to be.

\n

How much does it cost to run Dream-1? We want you to know exactly what to expect before you try it. Running the model costs just $3 per hour for Dream-1, or $2 per hour for Dream-1-flash. Each call to the model returns 1.6 seconds of actions, so an hour is 2,250 calls: about $0.0013 an action for Dream-1, and $0.0009 for Dream-1-flash.

\n

$3 / hour

\n

Dream-1

\n

Everything above the hourly cost is yours to keep.

\n

$2 / hour

\n

Dream-1-flash

\n

The faster, lighter model.

\n

Across tasks. We break Dream-1’s performance down into eight task categories, from what it handles well to where it still struggles.

\n

Fine-tune for your task with FaaS. If Dream-1 falls short on a task you care about, you don’t have to wait for our next release. Collect some data on your own robot, run it through FaaS, and get a custom model built for that task.

\n

We want anyone with a robot to work like a forward-deployed engineer: find a task someone needs done, collect a bit of data, adapt Dream-1 to it, and deploy. If people will pay for what your robot does, you keep everything above the hourly cost.

\n

Dream-1 is a research beta

\n

It will make mistakes, and it has not been validated for safety-critical use. Please test it in controlled settings, and take extra care whenever a robot is running near people. If you see unsafe behavior, please report it to safety@dreamlabs.ai.

\n

04.5

\n

One-shot embodiment learning

\n

When zero-shot isn't enough, we fine-tune Dream-1 for your robot. We call this Fine-tuning-as-a-Service (FaaS).

\n

Here's what that looks like in practice: timelapses of data collection on two of our partners' robots, 5 minutes each.

\n

Data collection timelapse

\n

Five minutes of data collection on two robots at 8x: reBot B601-DM on the left, Maker Arm on the right.

\n

After adapting on just those 5 minutes, the difference shows side-by-side with the zero-shot run. Dream-1 now understands a bit better the affordance and kinematics of the robot.

\n

reBot B601-DM

\n

Simulated twin

\n

Zero-shot

\n

Time0:00

\n

+ 5 min

\n

After FaaS

\n

Time0:00

\n

Maker Arm

\n

Real robot

\n

Zero-shot

\n

Time0:00

\n

+ 5 min

\n

After FaaS

\n

Time0:00

\n

Zero-shot and after FaaS on the same instruction, “go ahead and put the banana in the bowl”, with five minutes of the robot’s own data in between: reBot B601-DM in its simulated twin, Maker Arm on the real robot. 3x speed.

\n

Some embodiments sit much further from what Dream-1 already knows, and closing a gap that large takes more of the robot’s own data. Dexmate with Sharpa dexterous hands grasps with hands in place of grippers, and took 77 minutes of its data. On YAM, we swapped in different end effectors for this task, a compliant gripper and a modified knife gripper, and it took 98 minutes of data.

\n

After FaaS5x speed

\n

Dexmate, Sharpa dexterous hands, 77 minutes of data

\n

Pick bread from toaster and place on plate

\n

After FaaS5x speed

\n

YAM, compliant gripper and modified knife gripper, 98 minutes of data

\n

Cut the leaf off the strawberry

\n

After FaaS on embodiments far from Dream-1’s training data, on the real robots. 5x speed.

\n

Across twelve robots from our FaaS beta partners, fine-tuning lifts the language-following score on ten of them:

\n

Zero-shot versus after FaaS on the same twelve robots, in simulation, relative to YAM.

\n

One-shot adaptation still outperforms zero-shot, so a gap remains. Closing it is one of our main goals: the better zero-shot gets, the closer any new robot comes to working out of the box.

\n

05

\n

Looking forward

\n

Dream-1 today is a manipulation policy for bimanual robots with parallel grippers. It’s a starting point. We believe a strong, general robot foundation model can extend to mobile manipulators, dexterous hands, and bodies we haven't imagined yet, opening up many more of the things people actually want robots to do. Robotics won't be solved on day one, and Dream-1 has real limits, which we've tried to show transparently above.

\n

Robotics is still stuck in a difficult loop: every new robot or task demands new data and a custom policy, often with limited generalization. With Dream-1 as a robotics API, we want to break that loop and put robot intelligence into more people's hands, so everyday users, not only researchers, can discover new applications for general-purpose robot intelligence.

\n

Dream-1 . public API

\n

Put Dream‑1 on your robot

\n

If you're building robots or putting them to work on real tasks, we want Dream-1 to be the fastest way to make them useful. Tell us what you're working on.

\n

Use Dream-1 Today

\n

Dream-1 MicroDuck control

\n

A fine-tuned Dream-1 checkpoint controlling multiple embodiments, including the Microduck.

\n

Notes and references

\n
  1. 1MakerMods, Maker Arm: a bimanual arm setup from about $2k. ↩
  2. 2Yahoo Tech, Hugging Face's $399 Microduck robot has already sold more than 10,000 units, 1 September 2026: more than 10,000 units since pre-orders opened on 28 August. ↩
  3. 3DreamZero: the paper that introduced World Action Models, February 2026. ↩
  4. 4LAP also showed signs of life of embodiment generalization: Zha et al., LAP: Language-Action Pre-Training Enables Zero-shot Cross-Embodiment Transfer. ↩
  5. 5Chi et al., Universal Manipulation Interface: In-The-Wild Robot Teaching Without In-The-Wild Robots, RSS 2024: zero-shot transfer across arms that share one gripper, from wrist cameras alone. ↩
  6. 6Robocurve, a Public Benefit Corporation measuring and reporting frontier robotics capabilities to the public: Dream-1 on RoboDojo-RC Tier 1, October 18, 2026, CC BY 4.0. ↩
\n

Cite this post

\n
@misc{dreamlabs2026dream1,\n  title  = {Zero-Shot to New Robots: Introducing Dream-1},\n  author = {Dream Labs},\n  year   = {2026},\n  url    = {https://dreamlabs.ai/research/dream-1}\n}
","headings":[{"level":2,"text":"Breaking the chicken-and-egg problem","id":"breaking-the-chicken-and-egg-problem"},{"level":3,"text":"The Cambrian explosion of cheap, affordable robots","id":"the-cambrian-explosion-of-cheap-affordable-robots"},{"level":3,"text":"The result that started Dream Labs","id":"the-result-that-started-dream-labs"},{"level":2,"text":"Introducing Dream-1","id":"introducing-dream-1"},{"level":3,"text":"Get started in minutes","id":"get-started-in-minutes"},{"level":3,"text":"Zero-shot on robots it has never seen","id":"zero-shot-on-robots-it-has-never-seen"},{"level":3,"text":"How zero-shot embodiment generalization works","id":"how-zero-shot-embodiment-generalization-works"},{"level":3,"text":"What Dream-1 can and can't do","id":"what-dream-1-can-and-can-t-do"},{"level":3,"text":"One-shot embodiment learning","id":"one-shot-embodiment-learning"},{"level":2,"text":"Looking forward","id":"looking-forward"},{"level":2,"text":"Put Dream‑1 on your robot","id":"put-dream-1-on-your-robot"},{"level":2,"text":"Notes and references","id":"notes-and-references"},{"level":3,"text":"Cite this post","id":"cite-this-post"}]}}