Scrambler

The hand is entirely passive. The wrist and fingers are printed in TPU, so they absorb impacts and conform to the terrain, and the fingers are built on thin flexures that curl inward to grab whatever they touch. FEA was used to tune the wrist and finger geometry, trading compliance against the buckling limit. Removable micro-spikes adapt the same hand to rock, snow and ice. Nothing is motorized, on purpose: no new wiring, thermal management or power draw at sub-zero temperatures means fewer ways to fail.

Sensing works without electronics too. An AprilTag on every finger joint rotates as the finger flexes under load, and homography recovers each tag's pose from a single camera frame. From that, the finger's curvature is reconstructed and scored for how well it has engaged the surface, telling the robot whether a hold is strong or about to slip. It runs on the cameras the G1 already has, so the tactile feedback adds no failure points.

Reading the G1's URDF showed the real bottleneck: the wrist joints tolerate five times less force than the elbow. Compliance in the hand and forearm absorbs the contact and redirects the load path into the elbow. The robot was then trained on a 45° slope with domain-randomized ice, snow and rock, variable wind and friction, dropping onto all four limbs with the legs generating propulsion and the hands providing stability, ledge support and fall arrest.