Robotics practitioner: data is far from solved — zero-shot and real-hardware gaps remain
paigeinsf · x · 2026-11-01
Pushing back on claims that robotics data is solved or commoditized, a practitioner lists what such takes reveal: confusing egocentric data with all robot-training data; never testing robotics models zero-shot; never feeling the pain of overfitting; never deploying a policy on real hardware beyond sim; never testing generalization beyond training tasks or standard benchmarks like robodojo. Her conclusion: robotics data is far from solved.
More from Embodied
- ETH Zurich releases browser-playable Unitree G1 parkour policy with attention-based contacts — ChongZzZhang · 2026-11-01
- Two RTX 3090s for €1,700: How Do You Actually Run 72GB of VRAM? — AnouarRifi · 2026-11-01
- Meta reportedly building humanoid robot prototype with Chinese startup — Distinct-Question-16 · 2026-11-01
- Firefighting robot survives 30 minutes inside a 1,000°C furnace and keeps working — LexiLove · 2026-11-01
- Neuralink to start vision implants within months, aiming to restore sight for the blind — DimaZeniuk · 2026-11-01
- Archon Robotics, a Shanghai embodied-AI startup from UniAD and Huawei ADS veterans, wants to be China's OpenAI for humanoids — Scobleizer · 2026-11-01