Vision-Language Steering improves frozen robot policies on out-of-distribution tasks with no retraining
chris_j_paxton · x · 2026-10-01
- RoboPapers Episode 108 covers Vision-Language Steering (VLS), a method to improve generalization of pretrained vision-language-action (VLA) robot models.
- Pretrained VLAs often fail when tasks shift slightly out of distribution (e.g., obstacle position, visual appearance changes).
- VLS steers the sampling process of a frozen diffusion or flow-matching policy, adding better OOD robustness with no new training data.
More from Embodied
- IROS teardown is still all human labor: when will robots take over setup? — micoolcho · 2026-10-01
- Robot demo wows with tight reactive control, pouring sugar and stirring a glass — chris_j_paxton · 2026-10-01
- DIY robot U-BOT torque test: great on small hills, but overheats on the big one — _Stocko_ · 2026-10-01
- Nvidia open-sources Lyra 2.0, turning any image into an explorable 3D world — CurieuxExplorer · 2026-10-01
- Agentic HIL: an open-source MCP server that lets coding agents flash firmware and test real hardware — wyced21 · 2026-10-01
- NavHarness: training-free harness for lifelong embodied navigation, 83.7 s-SR on GOAT-Bench — Xunyi Zhao · 2026-10-01