Physical AI Notes
KO

Arms and hands: lessons

Turns the hand's target position and force into joint torques and contact. Also covers how the resulting errors affect data and simulators.

On the robot

No lessons are published yet. Here are the planned topics.

Overview

The loops running during one reach, how gear ratio determines an arm's character, why grasping is hard

  1. From policy output to the current loop: the loops that run during a single reachComing soon

    The cascade from policy through IK and the servo loop down to the current loop, showing where FK, IK, gear ratio and action chunks fit, with a brief look at the hand at the end.

  2. How gear ratio determines arm characteristics: backdrivability, reflected inertia and backlashComing soon

    How backdrivability, reflected inertia and backlash affect collision safety, teleoperation feel and the recorded data, without equations.

  3. Why grasping is hard: grippers, multi-fingered hands and tactile sensingComing soon

    Grasping framed as a problem of making contact and then holding on through friction.

Basics

FK, Jacobians and IK, joint sag and gravity compensation, grasping with a parallel gripper

  1. From joint angles to the end effector and back: FK, the Jacobian and IKComing soon

    Product-of-exponentials FK, τ=JᵀF, damped least-squares differential IK and singularities.

  2. Joints do not stop at the commanded angle: reproducing sag and compensating for gravityComing soon

    Verifying e=(τ_g−τ_ff)/kp and Δx≈L·e with kp and payload sweeps.

  3. Grasping with a parallel gripper: friction cones, gripper commands and current-based grasp checksComing soon

    Measuring grip force and slip, and judging grasp success, in MuJoCo and on the SO-101.

Advanced

Current loops, gear ratio and reflected inertia, impedance control and torque control, why IK is re-solved every cycle, the contract between actions and the controller

  1. From motor to joint: current loop, gear ratio, reflected inertia and heatComing soon

    From τ=n·η·K_t·i to reflected inertia and I²R heating, and how these map to simulator parameters.

  2. An arm that complies with external forces: impedance control and torque controlComing soon

    Task-space impedance, choosing stiffness, and the limits of arms with high gear ratios.

  3. Why IK is re-solved every cycle: redundancy, singularities and QPComing soon

    Null space, constraints and task priorities, extending to a fixed-base upper body.

  4. The contract between actions and the controller: how VLA outputs become torqueComing soon

    Action representations, normalization, interpolation and the offset between action and state.

In practice

Filling in the gap checklist with three measurements, deploying a policy on a real arm, classifying and automatically detecting grasp failures

  1. Does your arm have a low or high gear ratio: completing the gap checklist with three measurementsComing soon

    Measuring steady-state error under gravity, how that error changes with payload, and backdrive torque, then deciding which layer absorbs the gap (presets built from published specs only).

  2. The first week after deploying a policy on a real arm: timing, chunk boundaries and safety limitsComing soon

    Timestamp logging, observing chunk boundaries, driver limits and collision detection.

  3. Five ways grasps fail: failure taxonomy and automatic detectionComing soon

    Classifying failures into perception offset, collision, sag, slip and gripper timing, with current-based detection.