HKU Team's PhysicalRSI 1.0 Tops RoboDojo Benchmark at 1% of GPT-6 Astra's Cost

A research team at the University of Hong Kong’s MMLab has proposed a simple and efficient self-evolving embodied AI baseline, PhysicalRSI 1.0, which ranked first on the RoboDojo benchmark with a 31.38% average task success rate at roughly 1% of the token cost of GPT-6 Astra, the team said.

PhysicalRSI 1.0, short for Physical Recursive Self-Improvement, achieved a 36.27 score and 31.38% success rate under a unified five-category evaluation, compared with GPT-6 Astra’s 28.86 score and 22.21% success rate. The result represents a 39.6% relative improvement in success rate over Astra at a fraction of the cost. Astra reportedly consumes about 200,000 yuan in tokens for a full RoboDojo test run.

The system uses a “System 1–2 Co-Evolution” framework in which a general multimodal foundation model (“System 2”) analyzes task failures and rewrites the harness — the agent’s working program connecting tools, code, memory, and skills — while a separate execution layer (“System 1”) carries out physical actions. Candidate variants are evaluated in the physical environment, and better-performing versions are retained to produce the next generation. The foundation model’s weights remain frozen; what evolves is the harness configuration.

The approach addresses a key weakness of general-purpose models in robotic control: while Astra led overall, its precision-operation success rate was only 4%, versus 32.5% for PhysicalRSI. In key insertion, PhysicalRSI achieved 39.33% success across three seeds, compared with 0% for Astra, π0.5, and Simate-beta. In tube insertion, it reached 66% versus 0–22% for comparators. On a block-swapping task, PhysicalRSI scored 92.67% success versus Astra’s 18% and π0.5’s 0.67%. On a memory-dependent sequential cover-lifting task, it reached 100% versus Astra’s 46%.

In a cloth-folding iteration, success rate rose from 27.5% to 60% and score from 31.5 to 61.5 across five selected checkpoints, with different generations exploring retraction paths, wrist alignment, target selection, and cloth tracking.

The team positions PhysicalRSI 1.0 as a baseline and research infrastructure for embodied self-evolution, with unified interfaces for foundation models, VLA systems, tools, and physical environments, and a lineage-tracking system for candidate agents. The team said it plans to extend the framework to more robots and complex tasks.