Zero-Shot to New Robots: Introducing Dream-1

TL;DR

Dream Labs (ex-NVIDIA DreamZero team) comes out of stealth with $90M seed and launches **Dream-1**, the first public-API World Action Model (WAM) claiming *zero-shot embodiment generalization* to bimanual parallel-gripper robots never seen in training — including the 18-year-old PR2 — by predicting end-effector (EE) pose as the shared action space and offloading per-robot kinematics to each embodiment's own IK solver. Robocurve's independent evaluation reports Dream-1 above Claude Opus 5.5 and Gemini 4 Argon on RoboDojo-RC Tier 1, at 16× lower cost per task attempt than Gemini 4 Argon and 1.7× faster per call than GPT-6 Astra — framed as a new Pareto frontier for System-1 robot control at$3/hr (Dream-1) or $2/hr (Dream-1-flash). When zero-shot falls short, Fine-tuning-as-a-Service (FaaS) adapts the model with 5 minutes of robot data for close embodiments; wider gaps (Dexmate+Sharpa dexterous hands) take 77 min, modified-gripper YAM 98 min.

Key claims

• Dream-1 predicts EE-pose actions (per-arm 6-DoF pose + open/closed gripper) rather than joint angles, giving every bimanual parallel-gripper robot a shared action space that each embodiment’s own IK solver turns into joint commands [§04.3].
• Zero-shot embodiment generalization is enabled by the WAM substrate: because Dream-1 jointly predicts video and actions, controlling a robot and imagining what that robot will do are the same problem, and from the head camera every robot is just “shapes moving through a scene” [§04.3].
• UMI’s prior zero-shot cross-arm result was limited to policies that relied only on wrist cameras with identical gripper hardware; Dream-1 lifts both constraints — different gripper geometries and head-camera scene conditioning — by moving the invariance into the video prediction channel [§04.3].
• On Robocurve’s RoboDojo-RC Tier 1, Dream-1 reaches a Pareto frontier above Opus 5.5 and Gemini 4 Argon, at 16× lower cost per task attempt than Gemini 4 Argon and 1.7× faster per call than GPT-6 Astra; LLMs are good at planning, Dream-1 is pitched as the fast general System 1 that acts [§04.4].
• Dream-1 inference costs $3/hour and Dream-1-flash$2/hour, with each call returning 1.6 seconds of action — $0.0013/action for Dream-1,$0.0009/action for Flash [§04.4].
• One-shot embodiment learning via FaaS lifts language-following score on 10 of 12 fine-tuned beta-partner embodiments; closest embodiments need ~5 minutes of robot data, Dexmate+Sharpa dexterous hands took 77 minutes, and YAM with swapped compliant/knife grippers took 98 minutes [§04.5].
• Zero-shot performance depends partly on each robot’s IK solver, and the published performance breakdown is cross-robot in simulation only (not real) [§04.2, §04.4].
• The team cites DreamZero’s result that 30 minutes of YAM robot data was enough to generalize zero-shot to unseen tasks as the trigger for believing the chicken-and-egg data loop was breakable, with Dream-1 as the first public API around that recipe [§03].

Method

Dream-1 inherits DreamZero’s World Action Model architecture — a joint video + action predictor built on a pretrained video diffusion backbone, trained to predict future world state and future actions together. The architectural lever beyond DreamZero is the shared action-space interface: alongside embodiment-specific joint-angle outputs for natively-supported robots, Dream-1 emits per-arm 6-DoF EE pose + binary gripper state, which every bimanual parallel-gripper robot can execute via its own IK. Observations are three cameras: a head view plus two wrist views. The video-head channel is what carries the embodiment-invariance — the model’s “see shapes move, predict pixels” substrate is agnostic to whether the arm is a Franka, a YAM, or a 2008 PR2.

Deployment is a public REST-style API behind dreamlabs/dream-1 on the Python dream client. Fine-tuning-as-a-Service (FaaS) is the fallback when zero-shot falls short: collect robot-specific data on your embodiment, upload, receive a custom checkpoint. Dream-1-flash is a smaller/faster variant preserving the API shape at lower per-hour cost.

Results

• Zero-shot deployment reported on 20+ unseen bimanual parallel-gripper embodiments including the Willow Garage PR2 (released 2008) [§04.2].
• Robocurve independent evaluation (RoboDojo-RC Tier 1, 2026-10-18, CC BY 4.0) places Dream-1 on a new Pareto frontier: beats Opus 5.5 and Gemini 4 Argon on task success, at 16× lower cost per task attempt than Gemini 4 Argon, 1.7× faster per LLM call than GPT-6 Astra [§04.4].
• $3/hr (Dream-1) /$2/hr (Dream-1-flash) deployment cost, 1.6s of action per call → $0.0013/action and$0.0009/action respectively [§04.4].
• One-shot FaaS on 5 min of close-embodiment data closes most of the zero-shot gap; Dexmate+Sharpa dexterous hands at 77 min data and modified-gripper YAM at 98 min data close larger gaps; fine-tuning lifts language-following on 10 of 12 beta embodiments [§04.5].
• Dream Labs raised $90M seed; API publicly accessible [§opening].

Why it’s interesting

Dream-1 is the sharpest filed empirical datapoint for the “shared action space + WAM video channel” recipe for cross-embodiment generalization — it stakes out a cell on Human-to-Robot Retargeting‘s interface-location board where the invariance lives simultaneously in the EE-pose output and in the video-prediction supervision substrate, closest in spirit to Cross-Embodiment Robot Manipulation via a Unified Hand Action Space UHAS’s canonical fingertip-sphere interface and Transformer Transformer: A Unified Model for Motion-Conditioned Robot Co-design Transformer Transformer’s RoboToken-and-let-the-model-adapt framing but with a WAM backbone rather than per-task RL or diffusion-over-robot-tokens. Robocurve’s independent Pareto result above Opus 5.5 and Gemini 4 Argon is also the first filed external evaluation that pushes a dedicated System-1 robotics API ahead of frontier LLM controllers on both success and cost (sibling datapoint to Astranomy: An Empirical Study of GPT-6 Astra as a Native Robot Policy Astranomy’s Agentic Astra at 74.8% RoboLab-120 at $36/episode — Dream-1’s cost structure is two orders of magnitude lower). Companion to World Action Models are Zero-shot Policies (DreamZero) DreamZero: the research result (30-min YAM adaptation + zero-shot task generalization) made public API deployment plausible; Dream-1 is the productized next step with the FaaS loop closing the one-shot gap for new embodiments.

See also

• World Action Models are Zero-shot Policies (DreamZero) — predecessor WAM from the same team; established that 30 min of YAM data is enough for a WAM to generalize zero-shot to unseen tasks, which Dream-1 extends to zero-shot across unseen embodiments.
• Dyna-2: A 1-Million-Hour Scaling Law for World-Action Models — the other filed million-hour-scale WAM betting on human-video-to-robot transfer; contrasts with Dream-1 on data substrate (human video vs the DreamZero pretraining corpus) and on interface (wrist-EE + continuous grasp vs full 6-DoF EE-pose with shared video channel).
• Cross-Embodiment Robot Manipulation via a Unified Hand Action Space — shared-interface answer at the dexterous-hand level (canonical fingertip sphere); Dream-1 is the parallel-gripper analog at the EE-pose level.
• Transformer Transformer: A Unified Model for Motion-Conditioned Robot Co-design — “treat embodiment as decision variable” counter-recipe; Dream-1 treats embodiment as given and makes the model and action space invariant to it.
• Astranomy: An Empirical Study of GPT-6 Astra as a Native Robot Policy — Astranomy’s LLM-as-controller cost/latency numbers on RoboLab-120 ($36/episode); Robocurve’s Dream-1 numbers sit two orders of magnitude lower on the same Pareto axis.
• Inspect Robots — An Open-Source Evaluation Framework for Physical AI (Robocurve) — Robocurve’s open-source Inspect Robots harness, substrate behind the Dream-1 Pareto evaluation.
• VLA Models — Dream-1 is the sharpest “shared-action-space WAM” datapoint on the recipe-lever board.
• Human-to-Robot Retargeting — Dream-1 stakes out a new cell: shared-action-space (EE-pose) + shared-video-channel (WAM) + per-robot IK solver, with the invariance absorbed by both channels rather than one.