NVIDIA's Grana reward model boosts real-robot assembly success from 50% to 75%
yukez · x · 2026-11-21
NVIDIA, UT Austin and UW introduce Grana, a compositional reward model for long-horizon robot manipulation. A VLM estimates subtask progress and combines it with credit for completed steps, tracking progress lost to mistakes while preserving completed-step credit.
- Training-time weighted RL: on real-robot airplane assembly, Grana-weighted training hit 75% success vs 50% unweighted on the same data.
- Test-time action selection via MPC: the same policy went from 2/10 to 9/10 successes on tractor assembly.
Paper, code and video are public.
More from Embodied
- Tesla FSD 15's real-time view isn't 3D rendering — it's the AI model's own synthetic perception — mark_k · 2026-11-22
- China Installed 59% of New Industrial Robots in 2026 Report, ~354,000 Units — ingliguori · 2026-11-22
- Apple Watch bricks after one shower while a $15 Keep band survives many — oran_ge · 2026-11-22
- Tesla FSD V15 visualizations go viral with detailed 3D night-time rendering — DimaZeniuk · 2026-11-22
- Tesla FSD visualization: from dots on a screen to photorealistic 3D night views in six years — XFreeze · 2026-11-22
- YQ Bionics' Y-Hand M2 packs 38 degrees of freedom as AI meets robotic manipulation — CurieuxExplorer · 2026-11-22