Yacine shares trick: probe mingru hidden states to check domain randomization coverage
yacineMTB · x · 2026-11-05
Yacine MTB shared a technique for robot learning: train a probe to predict your simulator parameters from the mingru hidden state, letting you measure how well domain randomization captured the real world. He argues agents should be able to quantify how "seen" a state is, since exploration unlocks genuinely unseen experiences.
Related event: Drone Training Tips: Probing Hidden States to Validate Domain Randomization(3 posts)→
More from Embodied
- Robotics strategy: be the last mover and capture monopoly profits — StewartalsopIII · 2026-11-05
- An AI radio prototype from 2023 resurfaces: LangChain ReAct + GPT-4 in a Rams-inspired box — dee_hw · 2026-11-05
- Waymo driving itself on the highway at dawn: "will never get old, so dang sci-fi" — DynamicWebPaige · 2026-11-05
- DRAP paper proposes replay-free continual learning that cuts plasticity at dynamics-risk hotspots for physical AI — Oswaldo Ludwig · 2026-11-05
- Tiiny Pocket AI box tested: gpt-oss-120b hits 16 tok/s and is the only model to pass a 10-test coding benchmark — IronDiligent6698 · 2026-11-05
- Three robot posters and a world-model workshop talk at CoRL 2026 in Austin — breadli428 · 2026-11-05