Shengshu's Motus2 World Model Lets Robots Close the Loop on Self-Improvement

Shengshu Technology released Motus2 on September 12, a new world model for robotics that departs from the usual policy-model recipe of mapping the current observation straight to an action. Instead, Motus2 tries to do three things inside one model: decide what to do, predict what will happen after doing it, and judge whether the outcome was good, then use that judgment to revise its own policy.

Those three capabilities map onto three modules: a world action model (WAM) that generates actions, an action-conditioned world model (AC-WM) that predicts consequences, and a value model (VM) that scores outcomes. Together they form a chain of generate action, predict consequence, evaluate result, update policy. The predictions and scores the model produces feed back as signals for improving its own strategy, which then generates the next round of candidates, an arrangement QuantumBit describes as an initial exploration of recursive self-improvement.

For one model to both act and anticipate, it must not peek at the frames that follow an action before committing to it. Earlier video-generation-plus-action-control approaches often failed for exactly this reason, learning to reverse-engineer actions from future visual frames and looking smart on training data while breaking in the real world. Motus2 adopts an Action-first information flow from the mid-training stage onward, fixing the order as observation to action, then prediction, then evaluation.

The implementation splits across inference and training. At inference, Motus2 uses Best-of-N planning: it imagines several candidate actions and their consequences in parallel, has the value model score them, executes the best one, then re-observes the real world before the next round. At training time, those scores become policy-update signals, with higher-scoring candidates receiving stronger positive guidance, an approach the team calls model-based reinforcement learning. Only action-related parameters are updated while prediction and evaluation stay frozen, so feedback does not drag the already-learned model off course.

Real-robot results offer an early read on the loop. Across phone placement and multi-finger manipulation tasks, the base policy averaged 65% success; planning alone lifted that to 67.5%, model-based reinforcement learning alone to 72.5%, and the two combined to 75%, ten points above the baseline.

Beyond self-improvement, Motus2 adds touch and memory to vision, language and action. A lightweight tactile expert module reads the latest tactile feedback before a short action segment to refine it, reusing intermediate results already computed by the main model to keep feedback fast and cheap; on paper-cup extraction and paper tearing, adding touch raised average success from 60% to 72.5%. For memory, Motus2 caches recent real observations by default, and in hidden-block search and history-cued button tasks, retaining full history averaged 57.5% success, clearly ahead of compressing it.

On hardware, Motus2 has been deployed on high-degree-of-freedom hands including the 22-DoF single-hand Sharpa Wave and the 20-DoF WUJI Hand 2, demonstrating tasks such as screwing in a light bulb, turning book pages and multi-finger manipulation. It is trained on roughly 130,000 hours of human first-person (Ego) video plus hundreds of hours of robot trajectories and human-robot alignment data, in three stages: monocular Ego video pretraining, then binocular video with human motion data, then robot-domain adaptation. In a controlled comparison, a model pretrained only on human Ego data averaged 51% success across five real-robot tasks, while adding robot-domain mid-training pushed that to 84%, a 33-point gain.

Shengshu lays out a five-level path for general world models, from L1 generating worlds and L2 interacting with them, through L3 acting within them, to L4 autonomous world agents and L5 world organizers. By that scale, Motus2 already holds L3's core capability and is beginning to touch the autonomous decision-making, self-feedback and continual self-improvement that define L4.

The team is candid about the distance left to travel: tactile data still transfers poorly across robot bodies, and prediction reliability on longer tasks, long-term memory efficiency and continual learning in open environments all remain open problems. For embodied AI, the significance of Motus2 is less another round of success-rate gains than the arrival of evaluation and feedback as part of the world model's own loop.