NVIDIA's ACCRUE: continual off-policy RL cuts real-robot training time by up to 68.5%
nvidia · hf · 2026-10-15
Key points
- NVIDIA and KAIST propose ACCRUE, a continual off-policy RL framework for post-training robot foundation models (VLAs).
- It carries lightweight actor and critic networks across tasks: the inherited critic adapts only on current-task transitions and guides actor learning, prior behavior is preserved via distillation toward cached targets on a small set of retained episodes, and fixed visual features keep the learner compatible with base-policy updates.
- On contact-rich assembly tasks in simulation and real hardware, ACCRUE accelerates new-task learning with near-zero forgetting and cuts real-robot training time by 31.0%–68.5%.
More from Embodied
- Atlanta apartment complex deploys robot security dogs that patrol at night and talk to residents — 4KTV · 2026-10-15
- dm_robots co-founder lands in Hong Kong to build China network, starting two-week GBA tour — DominiqueCAPaul · 2026-10-15
- Rhoda's tactile-sensor robot retrieves objects blindfolded, sensing texture and heat — DhruvBatra_ · 2026-10-15
- Valeo and KE:SAI launch open-source initiative for generalist autonomous driving AI — abursuc · 2026-10-15
- ASENA: A Coding Agent Writes and Fixes Programs On-the-Fly to Drive a G1 Humanoid — Lianhuiq · 2026-10-15
- Bittensor SN44 bets on generative world models as a simulation engine for robotics — richdotca · 2026-10-15