ManifoldAI tops RoboDojo benchmark with in-context robot policy and recursive self-improvement
机器之心 · wechat · 2026-10-17
ManifoldAI's WorldScape R²SI took first place on the RoboDojo robot-learning benchmark, beating GPT-6-Astra and NVIDIA GR00T-N1.7.
Two core techniques
- In-context learning: WorldScapePolicy 2.5 learns new tasks from a single human demo without retraining — action transfer across object variants, multi-stage skill composition, and zero-shot long-horizon tasks via visual and event memory. It outperforms matched VLA baselines by multiples on unseen tasks.
- Recursive Robot Self-Improvement (R²SI): a two-layer agent system. The inner RobotAgent executes tasks via per-task skills (optionally calling the reactive model WSP); the outer RecursionAgent reviews execution logs across rounds, revises skills, or trains WSP, with updates auto-published after passing tests.
Cases: in pushT, a 723-step failed run led to a skill revision (decouple translation/rotation, staged gripper release) that scored full marks in 596 steps, later iterated to 389 (35% fewer); in packing, WSP is both invoked and continuously retrained by the outer layer.
More from Embodied
- Brain implants are restoring speech to people with ALS, Science reports — EricMTrautmann · 2026-10-18
- Meta wearables VP runs his AI agent across 3 phones and 23 pairs of glasses — nikola_mr64990 · 2026-10-18
- AMI-EV: spinning prism gives event cameras stable texture, Science Robotics 2024 — RexDouglass · 2026-10-18
- Neuralink patients control robotic arms, drones and games with thought alone — XFreeze · 2026-10-18
- Indie roboticist shows first neck actuation attempt for humanoid robot, AI-assisted design — _Stocko_ · 2026-10-18
- Retired Penn professor arrested for blocking Uber Eats delivery robot — johnnyApplePRNG · 2026-10-18