ShengShu Technology Introduces Motus2: A Cutting-Edge Self-Evolving Robot Model

ShengShu Technology Introduces Motus2: Revolutionizing Robotic Dexterous Manipulation

At the recent 2026 Inclusion Conference, Yihang Luo, the co-founder and CEO of ShengShu Technology, showcased a remarkable breakthrough in robotics—Motus2. This self-evolving general world model is designed specifically for robotic dexterous manipulation, uniting various functionalities into one cohesive system.

What Makes Motus2 Innovative?

Motus2 distinguishes itself by incorporating a unified model that encompasses action generation, consequence prediction, and outcome evaluation. By leveraging model-based reinforcement learning, it effectively translates predictions into actionable signals for policy enhancement. This innovative approach paves the way for continuous improvement in the manipulation capabilities of robotic systems.

The research indicates that Motus2 has achieved an impressive average success rate of 84% across five core real-robot tasks. These tasks include placing a ball, manipulating objects with multiple fingers, attaching an eraser, screwing in a light bulb, and positioning a phone accurately. Moreover, a separate policy-optimization study demonstrated how integrating model-based reinforcement learning with inference-time planning boosted the success rate from 65% to 75%. This 10 percentage point gain showcases the potential of sophisticated algorithms in enhancing robot performance.

Real-Life Applications and Demonstrations

Demonstrating the capabilities of Motus2, ShengShu Technology revealed real-robot scenarios involving various tasks, such as screwing in light bulbs, turning pages, opening cans, and locating hidden objects. These tasks require complex coordination and historical data to navigate spatial relationships. The dual-arm robot platforms utilized in these demonstrations featured advanced dexterous hands, including WUJI and Sharpa Wave technologies, highlighting the model's versatility and robustness.

A notable aspect of Motus2 is its ability to generate actions based on language instructions, robot states, and visual history through its policy interface. In tandem, a simulator interface predicts future visual outcomes, and the evaluator interface assesses task progression—this collaboration creates a seamless loop of action generation, consequence prediction, outcome evaluation, and resultant policy updates.

During task execution, Motus2 employs advanced planning techniques to generate multiple candidate actions, evaluate their effectiveness, and execute the most promising option. This adaptive model not only learns from successful actions but also incorporates fresh real-world observations to refine future decisions.

Training Methodology and Historical Context

To further enhance its capabilities, Motus2 relies on extensive training methodologies that incorporate large datasets gleaned from human interaction. The primary dataset consists of approximately 130,000 hours of recordings that provide foundational patterns of manipulation and object interactions. This broad training is complemented by robot-specific data, which tailors the learning experience to optimize the model's performance.

In scenarios where direct observation is challenging, Motus2 adeptly retains historical information. This feature, combined with a tactile expert system, allows the robot to adjust its actions based on real-time feedback, further refining its performance. For example, in tasks such as tearing paper or manipulating objects, invoking the tactile expert has improved success rates significantly—from 60% to 72.5%—demonstrating the model's adaptability and precision in complex tasks.