Delete object info from observations and PPO learns to search anyway — TU Darmstadt on its Humanoids 2026 paper
Jan_R_Peters · x · 2026-10-09
TU Darmstadt's IAS lab reshares its Humanoids 2026-accepted paper (same work as above) with a methodological takeaway:
- Interactive perception is learnable with end-to-end deep RL
- The "weird trick": remove object info from the policy's observations, and PPO still learns to search and manipulate using only joint angles
- One-liner: Sensing ≠ Perception — perception can emerge from pure proprioception
More from Embodied
- LLMs play Connect 4 via robot arms in MuJoCo; fable 5.1 wins at $17.40 in API calls — eigenron · 2026-10-09
- Prediction: frontier models as real-time robot policies at 10k-100k tok/s within 1-2 years — ATTlKA · 2026-10-09
- Hugging Face launches Robotic Episodes Viewer for 24k+ LeRobot datasets — mishig25 · 2026-10-09
- Blind humanoid walks, plays soccer and lifts suitcases with joint encoders only — accepted at Humanoids 2026 — Jan_R_Peters · 2026-10-09
- CARE certifies VLA inference speedups up to 10.8x with statistical guarantees — UMCP · 2026-10-09
- Corporate robot bets graded: only one binding order with unit counts found — lukas_m_ziegler · 2026-10-09