Recursia hits 28-53% zero-shot on LIBERO-Pro Long, far ahead of ASPIRE and π0.5
jiqizhixin · x · 2026-10-22
- Beijing Jiaotong University, Peking University, NYU, USC, and Beijing Lingxing Intelligence present Recursia, an Embodied Recursive Self-Improvement paradigm for compositional generalization.
- Results: zero-shot success on LIBERO-Pro Long under position/task perturbation is 28% and 53%; NVIDIA's ASPIRE scores 5.8 and 15.1 points lower, OpenVLA and π0 sit near 0%, and π0.5 manages just 8% and 1%.
- Core problem: even reliable individual skills (recognize, grasp, place) break when the second step happens in a world changed by the first — the bowl moved, the arm pose shifted.
- Method (figure): a shared executable stack across four layers (tools/skills/handoffs/verifications); an Executor logs failures on source tasks, a Proposer patches the right layer by failure class, and candidates must pass generality, mechanism, and paired regression checks before entering the shared stack. Target execution uses the frozen stack zero-shot, with task-direction success climbing over iterations.
Related event: Recursia: Embodied Recursive Self-Improvement Shines on LIBERO-Pro(2 posts)→
More from Embodied
- OpenTac open-sources full-stack 3D tactile sensing ecosystem for any robot surface — anand_bhattad · 2026-10-22
- Jensen Huang's GTC Berlin finale: one-prompt Hermes Agent builds and trains a robot in simulation — NousResearch · 2026-10-22
- Tesla Q3: FSD Supervised attach rate tops 75% of new NA deliveries, Semi enters volume production — elonmusk · 2026-10-22
- OpenTac team to demo tactile sensors at CoRL with small-scale sensor giveaway — anand_bhattad · 2026-10-22
- Kodiak Robotics to start humanless trucking 'within weeks,' already hauling for IKEA with safety drivers — Scobleizer · 2026-10-22
- RAX Uses a Frozen Coding Agent to Drive Robots: 96.7% on DexJoCo at 1/20 the Cost — du_yilun · 2026-10-22