Manifold AI says its WorldScape R²SI system topped RoboDojo’s Oct. 17 simulation leaderboard with an overall score of 46.58 on a 0-to-100 scale and a 39.35% in-budget success rate. Here, “in-budget” means completing a task within its official step limit. The project page compares R²SI’s score with 28.97 for a reference agent using the same GPT-6 Astra vision-language model. That is a comparison between full systems; it does not isolate how much of the difference came from R²SI’s skill revisions.
R²SI divides the work between two agents. The inner Robot Agent observes camera images, robot position data, the task and its past actions, then chooses a move through a Skill: a written instruction that tells it how to handle a task. The outer Recursion Agent reviews execution records, proposes changes to those instructions and tests each version before releasing it. When a task uses the reactive model, a Skill calls it to execute one chunk of motion; control then returns to the Robot Agent, which decides the next step from fresh observations. The outer agent can also train that model from execution records.
The project page gives two examples. In align_blocks, each candidate Skill was tested on the same eight layouts. The first version finished none within the 200-step cutoff; version four finished all eight within it. In push_T, the first attempt took 723 steps, above the 600-step limit; the revised Skill brought a run to 596, and a later version took 389. The frozen Skill then completed three randomized layouts.
R²SI covers 42 bimanual robot tasks and 54 variants. The project calls 46.58 its overall score on a 0-to-100 scale and says it aggregates the variants by task, but its page does not give the scoring formula. It also attributes part of the system’s performance to the wider harness, which can include both revised Skills and a trained reactive model; the same-model comparison does not show that the underlying GPT-6 Astra model itself improved.
These are project-reported RoboDojo simulation benchmark results, not a physical-robot trial. RoboDojo’s repository distinguishes its 42 simulation tasks from a separate set of 18 real-world tasks, and R²SI’s page identifies its result on the simulation leaderboard.
Illustrative photo by Nenad Stojković, “Robot arm works on a small component in a lab setting,” via Wikimedia Commons, licensed under CC BY 2.0. It does not depict R²SI or a RoboDojo evaluation.
Sources
- WorldScape R²SI project page, system design, score definition, benchmark scores and task examples.
- Official RoboDojo leaderboard, linked from the project page as the Oct. 17 ranking.
- RoboDojo official repository, benchmark scope and evaluation documentation.