Manifold AI's WorldScape R²SI self-improving robot agent tops the RoboDojo Agent Track at 46.58
The AMW Read
A lesser-known embodied-AI player posts a self-reported top score on a new benchmark using a self-improving agent design, which updates the embodied foundation-model debate but shows no proof of commercial traction.
Manifold AI's WorldScape R²SI self-improving robot agent tops the RoboDojo Agent Track at 46.58
Manifold AI (流形空间) says its WorldScape R²SI (Recursive Robot Self-Improvement) system reached a macro progress score of 46.58 and a 39.35% success rate on the Agent Track of RoboDojo. RoboDojo is a bimanual manipulation benchmark released in July and led by the University of Hong Kong's MMLab with Tsinghua, UC Berkeley and about twenty other institutions. GPT-6 Astra scored 28.97 on the same track. R²SI reached 15.5% on fine manipulation, 34.25% on long-horizon tasks and 35.83% on generalization. The system has two layers. An inner Robot Reactive Agent loads Skills and calls the company's WorldScape Policy (WSP) model. An outer Recursion Reflective Agent reviews past execution logs, rewrites Skills, and retrains WSP. Every update must pass tests before release. In one push_T task, a 723-step run that went over budget dropped to 389 steps through Skill revisions alone. Separately, WorldScape Policy 2.5 scales in-context-learning training data to about 100,000 hours.
This matters because embodied AI is moving away from collecting data and fine-tuning for each new task. Manifold AI is betting on two things instead: a robot that learns a new task from a single demonstration, and a robot that improves after each run. Its improvements are stored as readable, versioned Skills that can be rolled back, not hidden weight changes. When RoboDojo launched, the top simulation policy scored only 8.80% against 76.03% for human experts. That gap is why agent-system scores are now being watched as closely as raw policy quality.
For builders and investors, the open question is whether these benchmark gains hold up on real robots. The results are self-reported, and fine-manipulation success is still at 15.5%. Per the AI Market Watch index, the pipeline logged no Manifold AI items in the last 90 days versus one in the prior 90, though that count covers only name-matched, pipeline-ingested sources.
