Dream-1

Dream-1 is a robot-control model developed by Dream Labs. Its interface accepts camera images, the robot's current state, and a task, then returns actions for the robot to execute.^[1] Introduced in October 2026, it is a world action model for bimanual manipulation with parallel grippers.^[2] Dream Labs describes it as the team's first publicly accessible API model following its earlier DreamZero research.^[3]

How robot control works

Dream-1 jointly predicts video and actions from head and wrist cameras. End-effector poses and gripper commands transfer through the robot's inverse kinematics (IK) solver; native embodiments can use joint-angle output.^[2]

The hosted interface separates the model from the robot hardware. Dream Labs' homepage describes an observation-and-action workflow rather than a complete robot product: the operator supplies images, state, and an instruction; the service returns the actions. It also offers custom models obtained by uploading robot demonstrations.^[1]

Zero-shot transfer and adaptation

Dream Labs reports transfer to unfamiliar robot bodies without their demonstration data.^[2]

Transfer depends on IK quality. End-effector targets leave the arm's path unspecified; unfamiliar dynamics or joint limits can make predictions infeasible.^[2]

"One-shot embodiment learning" uses fine-tuning, not zero-shot transfer: five minutes for reBot or Maker Arm; 77 and 98 minutes for dexterous hands and modified YAM grippers, respectively.^[2] In the service workflow, Dream Labs trains on the uploaded demonstrations and supplies a custom model ID.^[1]

Evaluation

Simulation archive

The detailed simulation archive covers 27 unseen robot embodiments and the YAM training embodiment. It specifies 297 rollouts per embodiment: 147 language-following rollouts across 49 instructions, with three repetitions each, plus 150 different unseen tasks attempted once each. Graders assign progress values of 0, 0.25, 0.5, 0.75, or 1. The mean of those grades measures progress; the success rate counts only full completions.^[4]

The launch chart instead measures performance relative to YAM, normalized to 100%.^[2] The archive displays absolute results separately. For YAM, its rounded combined figures are 58% progress and 23% success over 297 rollouts. Its language-following results are 67% progress and 33% success over 147 rollouts; unseen-task results are 49% progress and 13% success over 150.^[4]

The archive's videos use 3x playback.^[4] Launch real-robot videos use 4x or 8x playback.^[2]

Third-party real-robot evaluation

Robocurve evaluated Dream-1 on its RoboDojo-RC Tier 1 benchmark before launch. All 120 Dream-1 trials in Robocurve's run export are dated October 18, 2026 (UTC) and used end-effector control through Dream Labs' hosted service; Robocurve dates Dream-1's public release to October 19, 2026.^[5][12] Dream Labs cited the result in its launch post and described Robocurve's evaluation as independent. The same post lists Robocurve among the beta-tester sites whose Dream-1 rollouts it shows.^[2]

As of November 1, 2026, Robocurve's leaderboard reports the following results, with 120 attempts per model configuration. Dream-1 has the highest mean score of the 12 configurations listed and lies on the Pareto frontier of both the cost and the time charts. The score measures average task progress, not the proportion of attempts fully completed.^[5][6]

The cost comparison uses Dream Labs' $3/hour hosted rate for Dream-1 and an RTX 5090 rental rate of $1.19/hour for MolmoAct2. The time measurements are not identical: LLM results use wall-clock attempt duration, while action policies that always run to their cap use the mean time required to reach their final score. Robocurve caps LLMs at 40 model calls and action policies at 3,600 control steps.^[5]

Robocurve's per-trial export gives 20 Dream-1 trials per task. Dream-1 fully completed 31 of its 120 attempts; the rest scored partial or zero progress.^[12]

RoboDojo-RC Tier 1 is Robocurve's adaptation of six real-world RoboDojo tasks, using I2RT YAM arms and the Inspect Robots evaluation framework. The evaluators selected the six tasks with the highest GPT-6 Astra scores from an earlier 18-task evaluation. They modified physical setups, instructions, and progress rubrics. Each trial is scored on observable task milestones, with 1 representing full completion. Robocurve's September 2026 report used human graders who knew which model they were scoring; in Robocurve's run export, all 120 Dream-1 trials are marked as autograded, and the 20 stand-up-bottles trials carry the label "Autograded (re-graded, old rubric)".^[6][12]

The tasks are packing and pouring fruit, stacking bowls, storing objects in a safe, capping a pen, standing bottles upright, and sorting objects. This six-task selection and its partial-progress grading define the scope of the reported result; they are not a measurement of all household work or performance on every robot.^[6]

Access, prices, and licensing

Launch prices were $3/hour for Dream-1 and $2/hour for Dream-1-flash: 1.6 seconds of actions per call, or 2,250 calls/hour.^[2]

The API terms describe usage-based fees deducted from prepaid credits. These inference charges do not describe the purchase price of a complete robot. The SDK is licensed under Apache 2.0, but it contains no model weights and performs no model inference locally. Dream Labs retains the model weights, including fine-tuned weights, and supplies hosted access instead.^[7]

A separate browser playground offers turns controlling real laboratory robots and a simulation. Its interface distinguishes queued prompts, recent history, and selected best runs.^[8]

Research-beta restrictions

The API Terms of Service, version 1.0 effective October 19, 2026, restrict ordinary access to research, evaluation, and development. Production or commercial deployment requires a separate written agreement. They also require people to remain outside the robot's hazard zone during operation unless separately agreed.^[7]

The service has no real-time delivery guarantee. Operators must supply independent local fail-safe controls, emergency-stop provisions, appropriate motion limits, and supervision. The terms describe an experimental beta and prohibit safety-critical use. Neither API availability nor a successful demonstration certifies a robot installation as safe.^[7]

Research background

Dream Labs identifies DreamZero as its team's preceding research.^[3] The February 2026 robot learning paper World Action Models are Zero-shot Policies describes joint video-and-action prediction. Its new-robot experiment used 30 minutes of play data for adaptation. That result concerned task generalization after embodiment adaptation, which differs from transferring to an embodiment without collecting any of its data.^[9]

The Universal Manipulation Interface, presented at Robotics: Science and Systems 2024, provides an earlier approach to transferring manipulation policies across robot platforms. It uses handheld grippers for demonstrations, wrist-camera observations, and relative action trajectories.^[10] Another preceding method, LAP, represents robot actions in natural language and reports transfer to unseen embodiments without embodiment-specific fine-tuning.^[11] These are related research approaches, not alternative names for Dream-1.

Continue reading

• WWorld action modelA world action model (WAM) is a robot policy design that builds action generation on a video world model backbone rather than on a vision-language model, so that a single network jointly predicts how a scene…Mentioned twice above
• RRobocurveRobocurve (Robocurve, Inc., which styles itself Robocurve PBC) is a San Francisco public benefit corporation that evaluates how well frontier AI models and robot-control models perform on real robots, and…Mentioned 11 times above
• RRoboDojoRoboDojo is a benchmark for generalist robot manipulation policies that combines a simulation suite of 42 tasks with a real-robot suite of 18 tasks, a standardized physical test platform called…Mentioned 3 times above
• RRobot learningRobot learning studies how robots acquire or improve behavior from data and experience.Also in Robotics
• UUniversal Manipulation InterfaceUniversal Manipulation Interface (UMI) is an open-source system for collecting robot manipulation training data with a handheld, camera-equipped gripper instead of an actual robot.Also in Robotics
• NVIDIA CosmosNVIDIA Cosmos is a world foundation model platform developed by NVIDIA for physical AI applications, including autonomous vehicles and robotics.Also in World Models

More to read

• V-JEPA 2
• Jeff Hawke
• Odyssey-3
• Mimic Robotics
• AR-WAM
• Lumo-2

Topics:RoboticsWorld Models

References

1. ^1 ^2 ^3Dream Labs. "Dream Labs". Official product homepage and interface description. Accessed October 20, 2026.
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Dream Labs. "Zero-Shot to New Robots: Introducing Dream-1". October 2026. Official release, control method, prices, adaptation examples, and limitations. Accessed October 20, 2026.
3. ^1 ^2Dream Labs. "Our Team". Official account of the relationship between DreamZero and Dream-1. Accessed October 20, 2026.
4. ^1 ^2 ^3Dream Labs. "Full sim eval rollouts". Evaluation archive, scoring method, embodiment selection, and YAM reference results. Accessed October 20, 2026. Archived copy (October 19, 2026); the original URL no longer resolves as of November 1, 2026.
5. ^1 ^2 ^3Robocurve. "Real-world evaluations of physical AI". RoboDojo-RC Tier 1 score, cost, time, and release-date tables. Accessed November 1, 2026.
6. ^1 ^2 ^3Sravanthi Machcha, Achu Menon, Sabrina Zou, Tzu Kit Chan, and Jay Chooi. "Opus 5.5 on RoboDojo-RC Tier 1: higher scores at lower cost". September 23, 2026. Benchmark construction, tasks, scoring, and evaluation limitations.
7. ^1 ^2 ^3Dream Labs, Inc. "Dreamlabs API Terms of Service". Version 1.0, effective October 19, 2026; sections 4.7, 5, 9.2, and 10. Accessed October 20, 2026.
8. ^Dream Labs. "Playground". Official real-robot and simulation interface. Accessed October 20, 2026.
9. ^Seonghyeon Ye et al. "World Action Models are Zero-shot Policies". arXiv:2602.15922, February 17, 2026. DOI: 10.48550/arXiv.2602.15922.
10. ^Cheng Chi, Zhenjia Xu, Chuer Pan, Eric Cousineau, Benjamin Burchfiel, Siyuan Feng, Russ Tedrake, and Shuran Song. "Universal Manipulation Interface: In-The-Wild Robot Teaching Without In-The-Wild Robots". Robotics: Science and Systems, 2024. Official project page.
11. ^Lihan Zha et al. "LAP: Language-Action Pre-Training Enables Zero-shot Cross-Embodiment Transfer". arXiv:2602.10556, February 11, 2026; revised February 15, 2026. DOI: 10.48550/arXiv.2602.10556.
12. ^1 ^2 ^3Robocurve. "All Robocurve runs". Run explorer and data export with per-trial dates, control mode, scores, outcomes, and grading method. Accessed November 1, 2026.

Show all 12 references