Manifold AI's WorldScape R²SI Tops RoboDojo Agent Track by Turning Robot Failures Into Skills
jiqizhixin · x · 2026-10-28
Manifold AI introduces WorldScape R²SI, a framework that turns robot execution experience into reusable improvements. An inner Robot Agent acts through task-specific skills; an outer Recursion Agent reviews execution records, revises those skills, and can retrain the reactive action model when a task uses it. All updates must pass tests before release, decoupling immediate task execution from slower cross-attempt system improvement.
Results: the framework reportedly ranks first on RoboDojo's Agent Track, ahead of listed GPT-6 Astra and GR00T-N1.7. In the pushT example, skill revisions cut an initially over-budget 723-step attempt to a successful 596 steps, then 389 on the same layout. After freezing the skill, three random layouts completed in 222, 312, and 256 steps — all within the 600-step limit.
More from Embodied
- Embodied AI startup DYNA hiring across robotics roles after $120M raise — JasonMa2020 · 2026-10-28
- Dyna Robotics shows commercial deployments: 200 picks/hr, 22-min hotel room turns, ≥97% pass rate — JasonMa2020 · 2026-10-28
- Physical Intelligence in talks to raise $1B+ at roughly $20B valuation, Bloomberg reports — rohanpaul_ai · 2026-10-28
- PaceVLA: per-chunk adaptive planning makes VLA robots up to 1.8x faster at same success — chris_j_paxton · 2026-10-28
- FUSE San Francisco robotics localization conference reveals lineup with Anduril CEO — kscottz · 2026-10-28
- Kinex physical coding agent self-evolves to solve all 365 RoboCasa tasks in 7 rounds — daibond_alpha · 2026-10-28