VLA drops to 0/10 in a new lab; code-based adaptation restores 10/10 across both
sherryyangML · x · 2026-09-19
The WetRobo team tested cross-lab generalization under changed physical setups: a VLA policy trained on demonstrations from Lab X scored 10/10 in Lab X but 0/10 in Lab Y.
After on-site adaptation, the coding-agent system achieved 10/10 in both labs — same task, different lab setup. The result suggests code-level adaptation can fix the cross-environment transfer failure that plagues VLA policies.
More from Embodied
- Sentdex pours cold water on 'rebuild Tesla FSD in an hour' demo: the hard part is perception — teortaxesTex · 2026-09-19
- GPT-Policy teaches robots from human video demos via in-context learning, no retraining — KyeGomezB · 2026-09-19
- First human vs robot fight: influencer Frankie Lepenna takes on Terminator robot tonight — cixliv · 2026-09-19
- Dev reverse-engineers Oura Ring 5 to stream accelerometer data and skip the subscription — JeremyNguyenPhD · 2026-09-19
- Developers are building a strawberry picking robot at Founders Inc — IsaiahBallah · 2026-09-19
- Simulating electro-permanent magnetic grippers in IsaacSim: 1360N holding force, Maxwell stress tensor — IsaiahBallah · 2026-09-19