Tactile Curiosity Drives Robot Grasping Without Task Rewards, CoRL 2026 Paper
pabbeel · x · 2026-10-20
A joint ETH Zurich and Berkeley paper, "Tactile Curiosity Drives Robot Interaction," accepted to CoRL 2026, shows how tactile feedback improves robot exploration. Key findings:
- RL from scratch: with no task reward, only agents curious about touch learn to grasp — every variant observes touch, so the difference comes purely from the curiosity component. The collected data is interaction-rich enough for fully offline pick-and-place learning.
- VLA post-training: adding touch as an observation helps a frozen VLA that never saw touch, but making the robot actively curious about touch works best, especially on contact-rich tasks.
Pieter Abbeel amplified the work.
Related event: ETH and Berkeley tactile curiosity paper accepted to CoRL 2026(2 posts)→
More from Embodied
- RealSense D457 stereo cameras now streaming on Intel robotics dev kit — chrismatthieu · 2026-10-20
- F1's 1978 fan car refutes Tesla wall-climber doubts: only control remains — theteknosaur · 2026-10-20
- Air Street's Benaich: Sereact's real traction beats robotics hype on X — nathanbenaich · 2026-10-20
- Berkeley's EyeRobot 2.0 Replaces Wrist Cameras with Active Gaze for Precise Manipulation — berkeley_ai · 2026-10-20
- NVIDIA joins Google Gemma Hardware Hackathon with $45K prizes and Jetson kits — agihouse_org · 2026-10-20
- Meme coin raised $361K to buy a Boston Dynamics robot dog whose job is spreading joy — adamamcbride · 2026-10-20