Whole-Body Tactile Representations for Motion Control
Benchmarking tactile representations for RL-based dynamic locomotion, deployed on an ANYmal-D quadruped with an FBG-based tactile skin. Semester project at the ETH Robotic Systems Lab (Spring 2026).
Semester project at the ETH Robotic Systems Lab (Spring 2026), supervised by Dr. Andrei Cramariuc and Dr. Robert Baines.
Learning-based controllers for legged robots perceive contact only indirectly, through proprioception and 3D perception. Whole-body tactile sensing would supply the contact signal directly — but the representation in which it should be fed to a learned policy is unresolved, and no benchmark existed to compare candidates.
The benchmark. I designed the dodging task, in which a quadruped must track an end-effector target with one foot while evading balls launched at its shank, scored by six metrics. Under one identical training recipe I compared proprioceptive-history baselines, an idealized single-point contact observation (FPB), and two taxel-grid representations matched to an FBG-based tactile skin: a single-cell map and TaxelKV, which derives a multi-cell activation from the ball–shank penetration geometry.
In simulation, the presence of tactile sensing matters far more than its representation: against the proprioceptive baseline, TaxelKV cuts the impulse per contact from 1.30 N·s to 0.45 N·s, the end-effector tracking error from 0.027 m to 0.011 m, and the clearance time from 0.26 s to 0.18 s.
On hardware, the representations separate sharply. Deployed on an ANYmal-D quadruped carrying the FBG-based tactile skin, the proprioceptive policy ignores shoves altogether; FPB responds but inherits the CNN’s localization error; TaxelKV, whose multi-cell activation stays closest to what the real sensor delivers, is robust to that single-point noise. The takeaway: closing the sim-to-real gap means modeling the representation on what the sensor delivers, not on the idealized quantities a simulator provides.
