Manifold AI's WorldScape R²SI Tops RoboDojo Agent Track With 46.58

RoboDojo was released in July by the University of Hong Kong's Multimedia Laboratory with Tsinghua University, UC Berkeley and nearly 20 institutions. It has 42 simulation tasks and 18 real-robot tasks across three dual-arm platforms, testing generalization, fine manipulation, long-horizon execution, memory and open semantic understanding. At release, the team evaluated 30 representative strategies; the top simulation average success rate was 8.80% and the top real-robot rate 12.8%, while human experts reached 76.03% and 100%. The report described the benchmark as the "Everest of embodied intelligence."

In the Agent Track, the benchmark evaluates whether a complete agent system can observe, decide and call tools from start to finish, not only a policy model's manipulation. GPT-6 Astra previously scored 28.97 on the track, with 4% fine-manipulation success, 8.25% long-horizon success and zero success on 16 tasks, according to the report. WorldScape R²SI reached 15.5% on fine manipulation, 34.25% on long-horizon tasks and 35.83% on generalization, the highest in the report's comparison.

The core framework, R²SI (Recursive Robot Self-Improvement), adds an execution-review-revalidation-self-improvement loop rather than replacing the robot's model. Its inner Robot Reactive Agent uses multi-view RGB images, robot state, task description and episode memory to make decisions and load skills. The WorldScape Policy is treated as a skill: it executes an action chunk and returns control for the next observation. The outer Recursion Reflective Agent reviews execution records across rounds, revises skills and trains or updates the policy. The report said updates must pass tests before automatic release and version management.

In the push_T task, where an arm must push a T-shaped object into a target area within 600 native action steps, the first run took 723 steps and was counted as unfinished. The outer agent identified incorrect grasping, coupled translation and rotation, and premature release; revised skills decoupled the movements and checked for offset in stages. A revised skill completed the task in 596 steps, and later iteration reduced the same layout to 389 steps, about a 35% cut. With the skill frozen, the robot completed three unseen layouts in 222, 312 and 256 steps, all within budget, the report said. In a packing task, the inner agent called the policy while the outer agent organized execution records into training data to update it.

The report linked the result to Manifold AI's in-context learning work. WorldScape Policy 2.0, released three months earlier, cut new-task setup to recording a demonstration, according to the report. It reached 75% success on the memory-dependent Shell Game, versus 30% for π0.5 and 50% for DreamZero. In cross-embodiment block stacking, it reached 60% with a target image and 70% with a video demonstration, while π0.5 and DreamZero reached at most 20%. The upgraded 2.5 expanded ICL training data to about 100,000 hours and focused on action transfer across objects, skill composition for long chains and visual and event memory for long-horizon tasks, the report said.