shengshu-ai/Motus2

What it solves

Motus2 addresses the lack of integration between world simulators and action output heads in embodied agents. Instead of simply appending an action head to a simulator, Motus2 creates a closed decision-and-learning loop where the agent can perceive, predict, act, evaluate, and improve its own performance in dexterous manipulation tasks.

How it works

Motus2 uses a single model with shared weights that provides three distinct control interfaces:

1. Policy (World-Action Model): Proposes candidate action chunks.
2. Simulator (Action-Conditioned World Model): Predicts the visual consequences of those actions.
3. Evaluator (Value Model): Assesses the predicted outcomes to determine if they are successful.

To improve, the model scales data from large-scale monocular egocentric data to synchronized stereo egocentric data, then adapts to the robot domain using robot trajectories and human-robot alignment data. It also incorporates tactile feedback for contact-aware control and uses a biomimetic platform with stereo vision and dual dexterous hands.

Who it’s for

Researchers and developers working on general-purpose embodied AI, robotic dexterous manipulation, and world models for robotics.

Highlights

• Self-Evolving Loop: Integrates policy, simulation, and evaluation into a single shared-weight model for continuous improvement.
• Data Scaling Path: Progresses from human egocentric video data to robot-specific trajectories.
• Multimodal Sensing: Combines stereo vision and tactile feedback for precise control.
• Biomimetic Hardware: Designed for a platform featuring dual arms and dual dexterous hands.