BlindTouch is a simulation-first tactile manipulation project. It trains and evaluates a three-finger robotic claw that handles unfamiliar objects using touch and proprioception alone, without camera input or object labels.
The project builds a complete MuJoCo and Gymnasium pipeline around a tactile claw, randomized contact-rich objects, behavior-cloned touch policies, tactile safety filtering, locked evaluation suites, and polished household-object showcase rendering.
- A custom MuJoCo claw model with three radially arranged fingers and 27 fingertip tactile cells.
- A Gymnasium environment that exposes only tactile pressure, joint state, and short observation history to the policy.
- Procedural object families covering round retention, asymmetric rigid contact, low-friction handling, fragile balance, and awkward poses.
- Skill teachers that generate successful touch trajectories for offline imitation learning.
- A hierarchical behavior-cloning policy with a touch-mode head and mode-conditioned continuous actions.
- A tactile safety filter that limits force spikes and protects fragile objects during rollout.
- Locked evaluation suites for repeatable policy selection and behavior comparison.
- A rendered household-object showcase with colored objects, tactile overlays, and success clips for presentation.
BlindTouch separates the problem into touch-only perception, mode selection, continuous control, and safety filtering. The current best model is a hierarchical behavior-cloned policy trained from successful tactile demonstrations. It predicts a manipulation mode such as probing, stabilizing, lifting, or releasing, then applies a mode-conditioned motor command for the fingers and palm.
The policy never receives privileged object identity, labels, or visual input. Household-style objects are used to visualize transfer behavior, while the training and evaluation pipeline remains grounded in randomized procedural objects.
BlindTouch includes PPO/SAC reinforcement-learning experiments, but the strongest current controller is a hierarchical behavior-cloned tactile policy with safety filtering.