Coding-agent on-site adaptation lifts new-lab success from 0/10 to 10/10
sherryyangML · x · 2026-09-19
The WetRobo team examined what happens when the physical setup changes across labs: a VLA trained on Lab X demonstrations 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 setup — code-level adaptation beats retraining for cross-lab deployment. During adaptation the agent also pulls in existing tools (SAM, OpenCV, NumPy, MuJoCo, RGB-D), producing solutions that compose several off-the-shelf components.
More from coding & agent
- GPT ported the classic shooter Soldat to the web as a playable multiplayer game — mayfer · 2026-09-19
- OpenClaw launches Multiplayer Mode to kill the 'meat proxy' in agent teamwork — heyneighbor · 2026-09-19
- Computer-use agents aren't GUI agents: best agents will code, call tools and use the CLI — DhruvBatra_ · 2026-09-19
- Muse agent books real flights end-to-end with zero errors, user reports — alexandr_wang · 2026-09-19
- Browser use is a commodity with no moat, argues dev as top sites may build native agent tools — tedddyoweh · 2026-09-19
- Open-source demo: TypeSafe Jev as a typed router inside a Flue agent on Cloudflare — irvinebroque · 2026-09-19