Back to portfolio

6-Axis Arm

Adaptive inverse kinematics that learns through hardware deformities

The Goal

I want an inverse kinematics stack that stays accurate even when the arm deforms over time or during deployment: backlash, wear, manufacturing tolerances, soft mounts. If the controller can absorb those imperfections online, everyday IK applications can ship with much looser (cheaper) mechanical tolerances.

Hebbian Control Policy

The plan centers on a partial Hebbian network for the control policy. Unlike most reinforcement learning setups that freeze after training, Hebbian plasticity keeps learning from live hardware. That means the policy can train on deformed joint data and make decisions that already account for those deformities.

In practice this sits on top of a model-based operational-space tracker, with the adaptive layer correcting for unmodeled error so the arm keeps tracking paths as the plant drifts.

Arm tracking a path in MuJoCo with the adapter on (DR 0.4)
End-effector error: OSC alone vs OSC + Hebbian adapter under domain randomization

Why It Matters

Classical IK assumes the kinematic model matches the robot. Real arms diverge. An always-learning adapter turns that mismatch into something the controller can keep eating, instead of something that forces a recalibration ritual or tighter (and more expensive) machining.

Back to all projects