Berkeley researchers introduce Advantage Guidance, an RL recipe for VLAs hitting ~70% success
berkeley_ai · x · 2026-10-21
UC Berkeley's Zhiyuan Zhou and Sergey Levine introduce Advantage Guidance, an improved RL recipe building on the advantage-conditioned VLA training the author did for π0.6 during his Physical Intelligence internship.
- Core idea: stable advantage-conditioned supervised training plus contrastive guidance sampling, turning off-policy RL into a simple supervised learning loss
- Benchmark charts show AG (with contrastive guidance) reaches roughly 70% aggregated success rate, beating EDP, QGF, CFGRL, CRR, FQL and other RL baselines
- Paper and code are both available
More from Embodied
- Phonon TTS pipeline ported to GenieX, targeting on-device NPU inference — HowDevelop · 2026-10-21
- Ropedia releases open W0 world model trained on 10,000 hours of egocentric human video — liuziwei7 · 2026-10-21
- Characterizing a cheap SG90 servo with Claude's plan to build a MuJoCo model — _Stocko_ · 2026-10-21
- The New Yorker: We're Learning to Live With Waymos, and It's Going Weird — njrc · 2026-10-21
- Autonomous aviation startup Reliable Robotics raises $115M, topping $1B valuation — Sethwinterroth · 2026-10-21
- GPT-6 Astra controls a robot arm to garnish water with lemon, dinner party test begins — mmmbchang · 2026-10-21