ShengShu Technology co-founder and CEO Yihang Luo introduced Motus2, a general world model for robotic dexterous manipulation that the company describes as self-evolving, at the 2026 Inclusion Conference on the Bund on September 10. The company said Motus2 combines action generation, consequence prediction and outcome evaluation in a single video-action model with shared parameters, and applies model-based reinforcement learning so that predictions and evaluations of action outcomes act as signals for improving its policy.

According to the team's research paper, Motus2 reached an average success rate of 84% across five primary real-robot tasks: placing a ball, multi-finger manipulation, attaching an eraser, screwing in a light bulb and placing a phone. In a separate policy-optimization study involving phone placement and multi-finger manipulation, combining model-based reinforcement learning with inference-time planning raised the average success rate from 65% to 75%. These real-robot evaluations used 20 trials per task for each method.

The company said the model works through three interfaces. A policy interface generates candidate actions from language instructions, robot states and visual history; a simulator interface predicts the future visual states those actions may produce; and an evaluator interface judges whether the predicted outcomes would advance the task. During execution, Motus2 uses Best-of-N planning to generate several candidate actions, compare their predicted outcomes and run the higher-scoring option, replanning with fresh observations from the real world after each action chunk. During training, value scores assigned to candidate actions update the policy, and only action-related parameters are updated while the prediction and evaluation components stay frozen. The company said successful demonstrations provide targets for action learning, while outcomes from failed and suboptimal interactions help the model learn action consequences and assess task progress.

Training draws on an egocentric human dataset of about 130,000 hours of raw recordings from monocular and stereo sources, the company said. The model first learns broad object-change and manipulation patterns from monocular video, then uses stereo video and human action data to learn spatial relationships and hand-object interactions. Robot-domain mid-training uses more than 100 hours of robot trajectories and supplementary human-robot alignment data. Under the same target-task fine-tuning procedure, a model pretrained only on egocentric human data averaged 51% success across the five primary tasks, rising to 84% when robot-domain mid-training was added.

Motus2 also investigates mechanisms for retaining historical information and incorporates a lightweight tactile expert that reads contact feedback immediately before a short action segment is executed and refines the action, according to the company. Across two real-robot tasks, pulling out a paper cup and tearing paper, adding the tactile expert increased the average success rate from 60% to 72.5%. Within ShengShu Technology's five-level roadmap for general world models, Motus2 implements L3, “Acting in the World,” and explores a technical path toward L4, “Autonomous World Agents.” The company said the model architecture, research paper and real-robot demonstrations are publicly available, and that reliable decision-making over longer tasks, retention of important historical information and long-term autonomous learning in open environments remain areas for further research. ShengShu Technology was founded in 2023.