Shengshu Technology Releases Motus2 World Model for Robot Self-Improvement

Motus2's core change is self-evolution. Traditional robot policy models learn statistical links between observations and actions, such as seeing A and doing B, without understanding why B leads to the desired result C, the report says. Motus2 places three functions in one model: a world action model that generates actions, an action-conditioned world model that predicts consequences, and a value model that evaluates results. These functions form a loop: generate an action, predict its outcome, evaluate the outcome, and update the policy. Shengshu says this self-evolution means using the model's own prediction and value feedback to improve policies, not that a robot can learn indefinitely on its own in open environments.

To avoid cheating, the model uses an action-first information flow from the mid-training stage, so it first generates an action from current observations, then predicts the result, and finally evaluates it. At inference, Motus2 uses Best-of-N planning: it proposes several candidate actions, imagines their outcomes, scores them with the value model, and executes the highest-scoring one before re-observing the real world. During training, those scores become policy-update signals under what the team calls model-based reinforcement learning. Only action-related parameters are updated, while prediction and evaluation remain fixed to avoid destabilizing the model. Failed trajectories also become useful, because success data show positive solutions while failures help the model learn action consequences and outcome evaluation. Physical Intelligence's RECAP also lets robots learn from practice and human corrections, the report says, while Motus2 uses a world model to predict and compare candidate actions without trying each one in reality.

In real-robot tests involving phone placement and multi-finger manipulation, the base policy had an average success rate of 65%. Planning alone raised it to 67.5%, model-based reinforcement learning alone to 72.5%, and the two together to 75%, 10 percentage points above the base policy, according to the report.

Motus2 adds touch and memory. A lightweight tactile expert module reads the latest tactile feedback before short action segments and refines the action, reusing intermediate results from the main model to balance feedback frequency and computation. In two real-robot tasks, extracting paper cups and tearing paper, touch raised average success from 60% to 72.5%, a 12.5 percentage-point gain. For memory, the model caches recent real observations. In tests involving finding a hidden block and operating buttons based on historical cues, keeping full history achieved an average success rate of 57.5%, clearly above compressed-history approaches. The report says the result does not prove memory compression cannot work; preserving full observations is an interim compromise before a better answer is found.

The model has been deployed on high-degree-of-freedom hands including the 22-degree-of-freedom Sharpa Wave and the 20-degree-of-freedom WUJI Hand 2, demonstrating tasks such as screwing in light bulbs, flipping book pages and multi-finger manipulation. Motus2's data strategy relies on human first-person Ego data and robot data. Training proceeds in three stages: pretraining on large-scale monocular Ego videos to learn basic object and scene changes; adding binocular video and human motion data to learn finer hand-object interaction; and adapting to robot control spaces using robot trajectories and human-robot alignment data. The Ego data system has about 130,000 hours of human first-person data plus hundreds of hours of robot and human-robot alignment data. In one comparison, a model pretrained only on human Ego data reached 51% average success on five real-robot tasks; adding robot-domain mid-training raised it to 84%, a 33 percentage-point increase. Increasing binocular data from 2,000 hours to 20,000 hours continued to lower validation error on human motion prediction without an obvious ceiling, the report says.

Shengshu positions Motus2 within a five-level roadmap for general world models: L1 generates a world, L2 interacts with it, L3 acts in it, L4 is an autonomous world agent, and L5 is a world organizer. Motus2 has core L3 capabilities and begins to touch L4 through autonomous decision-making, self-feedback and continuous self-improvement, according to the report. The company and the report note that challenges remain, including transferring tactile data across robot bodies, reliable prediction over longer tasks, efficient long-term memory and continuous learning in open worlds.