Contact-rich manipulation: in-hand reorientation, fine grasping, tool use, handling deformables. It requires high-frequency closed-loop control fusing vision and touch under contact dynamics that are hard to sense, model, and simulate. The bottleneck for useful robots. Locomotion is largely solved; hands are not. ~85% confidence this is the gating problem for general-purpose robots.
Why it’s the bottleneck (five coupled reasons)
- Sensing — robots lack the dense, fast, multimodal touch humans have (Tactile Sensing & Electronic Skin); hands move faster than current tactile readout can process.
- Control — high DOF + stiff contact dynamics make control/RL hard (needs the ~200Hz System-1 policies from Vision-Language-Action (VLA) Models).
- Simulation — can’t cheaply generate contact-rich training data (Sim-to-Real, Robot Simulation & Synthetic Data; 5,000+ real demos for deformables).
- Data — teleoperating a 20+ DOF hand is itself hard (embodiment gap, operator fatigue), so demonstration data is scarce.
- Hardware — the anthropomorphic/durable/cheap/sensorised four-way trade-off is unsolved (Robot Actuators (the muscle layer)).
This is the mechanism behind Robot Autonomy Destination: the data bottleneck is specifically a contact/tactile bottleneck, which is why video-pretraining (good for “what to do”) doesn’t fix the physical “how.”
Hand landscape (DOF / actuation / cost / tactile)
- Shadow Dexterous Hand — ~24 DOF, tendon-driven, fingertip tactile options. Research gold standard, ~£100k+ class.
- Wonik Allegro V4/V5 — 16 DOF, direct-drive, ~$17k. The affordable research workhorse for in-hand RL.
- PSYONIC Ability Hand — prosthesis-origin, ~6 motors, fingertip touch, low cost.
- Clone “Hand of Clone” — 27 DOF, hydraulic myofiber muscles. Most biomimetic, the wildcard (production claims unproven).
- Tesla Optimus Gen3 — 22 DOF/hand, ~25 actuators relocated to the forearm (the notable engineering move).
- 2025 wave (mostly Chinese: TESOLLO, Unitree Dex5, ROBOTERA, PaXini) — 12-20 DOF, mostly direct-drive, all shipping with tactile. Rapid cost-down.
Connections
Physical AI (robotics cluster hub) · Tactile Sensing & Electronic Skin · Robot Actuators (the muscle layer) · Robot Autonomy Destination · Tactile Sensing Silicon · Humanoid Robots