Robots Atlas>ROBOTS ATLAS
Dream-1
AI Models

Dream-1

1
World Action Model by Dream Labs that controls bimanual, parallel-gripper robots through a public API, generalizing zero-shot to unseen robots, tasks and environments.
🧪 Beta✓ Public accessWorld ModelRobotics foundation model
🏢Dream LabsProducer
Access:APIHostedDeployment:☁ Cloud

Overview

Dream-1 is a World Action Model (WAM) built by Dream Labs in San Francisco, released in October 2026 alongside the company coming out of stealth with a $90M seed round. It controls bimanual robots equipped with parallel grippers and is accessed through a public API.

Unlike policies that output joint angles, Dream-1 predicts end-effector (EEF) poses: where each gripper should be and whether it is open or closed. This shared EEF action space lets it command robots it has never seen, while the robot’s own inverse-kinematics (IK) solver turns the poses into joint commands.

As a WAM, Dream-1 learns to predict video and actions jointly — scenes from three cameras (a head view and a wrist camera on each arm) become a shared representation that lets the model generalize zero-shot to unseen robots, tasks and environments. It builds on DreamZero, the February 2026 paper that introduced World Action Models.

The model ships in Dream-1 and a lighter, faster Dream-1-flash variant; Dream Labs also offers per-robot adaptation via Fine-tuning-as-a-Service (FaaS) from a few minutes of the robot’s own data. Dream-1 is described as a research beta and has not been validated for safety-critical use.

Classification
World ModelRobotics foundation model
Access & deployment
APIHosted
Cloud
Weights: Closed
Key parameters
✓ Fine-tuning
📥 Input: image, text, robot state data
Robotics
Bimanual manipulationRobot manipulationRobot controlObject affordance understanding

Technical specification

Hardware requirements
A bimanual robot with parallel grippers and three cameras (a head view plus a wrist camera on each arm). Runs via Dream Labs’ cloud API and relies on the robot’s own inverse-kinematics (IK) solver.
Features:✓ Fine-tuning
Modalities
⬇ Input
imagetextrobot_state_data
⬆ Output
robot_actionsmanipulator_controlmotion_trajectories

Capabilities and applications

Native model capabilities
Zero-shot learning
The model's ability to perform a new task without dataset-specific training or hyperparameter tuning — prediction is produced in a single pass from context.
Category: other
Cross-embodiment transfer
The ability of a single model to control robots with different morphologies (humanoids, dual-arm rigs, mobile platforms) without training a separate model per platform. Intelligence is decoupled from embodiment, so the same policy runs on hardware with different kinematics and dynamics.
Category: robotics
Vision-language-action grounding
The ability of a VLA model to ground visual perception and a language instruction into a concrete physical robot action. The model understands the scene and intent, then generates an executable action sequence, closing the loop from observation to motion.
Category: robotics
Action conditioning
Controlling model generation via action signals (camera, robot pose, commands, speech) rather than text prompts alone.
Category: multimodal
Instruction following
Precisely following instructions contained in the prompt: response format, length, style, constraints (e.g. 'reply in six words'). GPT-5.1 significantly improved this capability compared to GPT-5.
Category: language
Real-time inference
The model's ability to generate responses with very low latency (>1000 tokens/sec) on specialized inference hardware (e.g. Cerebras WSE), enabling interactive, turn-by-turn collaboration with a human.
Category: coding
World simulation
Model's ability to generate coherent, interactive simulations of physical environments — maintaining geometry, lighting, and physics during exploration.
Category: multimodal
Few-shot learning
The model's ability to perform a new task from a handful of examples provided directly in the prompt, without any weight updates or fine-tuning.
Category: other
Robotics
Bimanual manipulationRobot manipulationRobot controlObject affordance understanding

Benchmark results

1 benchmark
RoboDojo-RC Tier 1
mean score per task attempt · autonomous rollouts on bimanual robots
📅 18 Oct 2026📄 Robocurve (Public Benefit Corporation) – niezależny raport
Robocurve’s independent report places Dream-1 above Opus 5.5 and Gemini 4 Argon, at 16x lower cost per task attempt than Gemini 4 Argon and 1.7x faster than GPT-6 Astra.

Pricing

Technical architecture

Sources and related pages