Dexterous Manipulation & Robot Hands

last updated 2026-06-13
Tactile Sensing & Electronic SkinTactile Sensing & E…Robot Actuators (the muscle layer)Robot Actuators (th…Vision-Language-Action (VLA) ModelsVision-Language-Act…Sim-to-Real, Robot Simulation & Synthetic DataSim-to-Real, Robot …Dexterous…

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)

  1. Sensing — robots lack the dense, fast, multimodal touch humans have (Tactile Sensing & Electronic Skin); hands move faster than current tactile readout can process.
  2. Control — high DOF + stiff contact dynamics make control/RL hard (needs the ~200Hz System-1 policies from Vision-Language-Action (VLA) Models).
  3. Simulation — can’t cheaply generate contact-rich training data (Sim-to-Real, Robot Simulation & Synthetic Data; 5,000+ real demos for deformables).
  4. Data — teleoperating a 20+ DOF hand is itself hard (embodiment gap, operator fatigue), so demonstration data is scarce.
  5. 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)

Connections

Physical AI (robotics cluster hub) · Tactile Sensing & Electronic Skin · Robot Actuators (the muscle layer) · Robot Autonomy Destination · Tactile Sensing Silicon · Humanoid Robots

Recent mentions

Related concepts

Frontier questions