Kinesthetic teaching beats teleop for dynamic and precise SO-101 tasks, but can't record grip force
kamathsblog · x · 2026-11-13
Developer giacomoran stumbled into kinesthetic teaching while building UMI for the SO-101 (5 DoF):
- Throwing a ball: far more natural than teleoperation for dynamic tasks.
- Playing a GBA: precise enough to press tiny buttons, which standard SO-101 teleop struggles with badly.
- Picking up a cube: reveals the key flaw — rigid fingers only record position, not grip force. In teleop, the follower's servo converts the leader's over-travel into squeeze force, so the policy learns to command the gap. With kinesthetic teaching, touching and squeezing the cube give identical positions, so the policy never learns to actually grip.
Takeaway: teleop records grip force almost by accident; soft fingers or force sensing may be needed to close the gap.
More from Embodied
- EmbodiedSWE wins CoRL 2026 Best Paper: coding agents generate robot training data — PeterHndrsn · 2026-11-14
- Ducted-Fan Robot Uses Coandă Effect for 3D Thrust Vectoring and Stable Flight — bennash · 2026-11-14
- ASUS teases its first humanoid robot in near-black teaser video — Distinct-Question-16 · 2026-11-13
- EmbodiedSWE, coding agents for dexterous robotics, wins Best Paper at CoRL 2026 workshop — shahdhruv_ · 2026-11-13
- Kodiak's in-house sim stack BreakPoint cuts costs to 2% while running 16x faster — Scobleizer · 2026-11-13
- YC startup Antimatter builds voice-first hardware after AirPods misheard 30% of words — aakashgupta · 2026-11-13