Shengshu Technology's Motus2 World Model Lets Robots Self-Improve Through Feedback

Motus2 follows Motus, released in February when world action models were still an immature concept. QbitAI said Motus beat Pi-0.5 by more than 35 percentage points in absolute success rate on 50 general task tests. Motus2 expands vision, language, action and touch modalities and combines three capabilities in one model: acting, predicting and evaluating.

The core change is a closed loop. A world action model generates actions, answering what to do next; an action-conditioned world model predicts consequences, answering what will happen after the action; and a value model evaluates results, answering whether the result is good. The loop is generate action, predict consequence, evaluate result, update policy. The model's own predicted and evaluated results become feedback for improving its policy, and the improved policy enters the next round.

To prevent the model from peeking at future outcomes before acting, Motus2 uses an action-first information flow starting in robot-domain mid-training. At inference, it uses Best-of-N planning: it proposes several candidate actions, imagines their outcomes, has the value model score them, and executes the highest-scoring one; after a short segment, the robot observes the real world again and starts another round. At training, those scores become signals for policy updates, a method the team calls model-based reinforcement learning, or MBRL. Only action-related parameters are updated, while prediction and evaluation parts remain fixed.

In two real-robot tasks, phone placement and multi-finger manipulation, the base policy averaged 65 percent success. Planning alone raised it to 67.5 percent; MBRL alone reached 72.5 percent; combining both reached 75 percent, 10 percentage points above the base policy. The closed loop also changes data use. Past work relied heavily on perfect trajectories and filtered out failures as noise. In Motus2, successful trajectories show positive solutions, while failed trajectories help the model learn action consequences and result evaluation. QbitAI noted that Physical Intelligence's RECAP also lets robots learn from humans and practice, with human takeover when they err, but Motus2 uses the world model so some candidate actions can be predicted and compared before being tried in reality.

Motus2 adds touch and memory. In dexterous manipulation, vision can see where an object is and its pose, but it struggles to fully judge contact. When tearing kitchen paper, correct-looking hand positions do not guarantee the paper is firmly gripped; slipping or force changes affect the next action. Motus2 uses a lightweight tactile expert module that reads the latest tactile feedback and refines actions before a short action segment is executed. It reuses intermediate results from the main model instead of rerunning the full video backbone each time. In extracting a paper cup and tearing paper, adding touch raised average success from 60 percent to 72.5 percent, an increase of 12.5 percentage points.

Memory addresses tasks where a single current frame is insufficient. In one test, a block is hidden in one of three cups and the cups are rearranged, so the robot must remember what happened. Motus2 by default caches recent real observations. In tests of finding a hidden block and operating a button based on a historical prompt, retaining complete historical information averaged 57.5 percent success, above a scheme that compressed historical information. QbitAI said this does not prove memory compression is unworkable; saving complete observations is more like a temporary compromise. Other teams keep only a recent short observation or highlight history by keeping the beginning and recent frames while compressing earlier steps.

Motus2 has been deployed on high-degree-of-freedom dexterous hands including the 22-DOF single-hand Sharpa Wave and the 20-DOF single-hand WUJI Hand 2, demonstrating tasks such as screwing in a light bulb, turning book pages and multi-finger manipulation. More degrees of freedom create more possibilities but also make learning complex, and the real difficulty is data. Motus2 treats human first-person, or ego, operation experience as a main data source and gradually transfers it to robots. Nvidia's EgoScale uses a similar idea. Motus2's multi-level ego data system has three training steps: monocular ego video pretraining to learn basic object and scene change patterns; introducing binocular video and human motion data to learn finer-grained hand-object interaction; and robot-domain adaptation using robot trajectories and human-robot alignment data to transfer earlier steps into the robot's control space.

The system uses about 130,000 hours of human first-person data plus hundreds of hours of robot and human-robot alignment data. In a controlled experiment, under the same target-task fine-tuning process, a model pretrained only on human ego data averaged 51 percent success on five real tasks; adding robot-domain mid-training raised success to 84 percent, an increase of 33 percentage points. A separate data-scale experiment increased binocular data from 2,000 hours to 20,000 hours and found validation error on human motion prediction continued to fall without a clear ceiling, suggesting that scaling human data has not yet reached its limit.

Shengshu Technology divides its general world model evolution into five levels: L1 generate the world, L2 interact with the world, L3 act in the world, L4 autonomous world agent, and L5 world organizer. These are not independent model types but an accumulating capability path. By this scale, Motus2 already has core L3 capabilities: understanding current state, predicting action consequences and outputting real robot actions. Moving from L3 toward L4 raises a key question: can the model use its own abilities to help improve itself? This echoes RSI. In Motus2, the model predicts consequences of different actions and uses value evaluation to produce policy update signals; the updated policy continues to generate candidate actions and enters the next round of prediction, evaluation and optimization.

QbitAI reported that the work still faces limits: tactile data remains difficult to transfer across robot bodies, and long-horizon prediction reliability, long-term memory efficiency and continual learning in open worlds all need further exploration. Still, Motus2 validates at least one point: from predicting action consequences to using those consequences to improve policy, evaluation and feedback have started to enter the closed loop of a world model.