Software & Machine Learning

Bimanual Arm Inverse Kinematics

Jacobian-based iterative IK for the dual-arm setup, driving both end effectors to target poses in simulation.

6 DOF Arm + 15 DOF Hand Jacobian Iterative Based IK Controller

Inverse kinematics: fingertip_ik.py

We implemented iterative Jacobian IK with damped least squares (DLS) to take in the 5 target fingertip position and outputting the joint angle that will make it happen. It utilize the Jacobians and body positions from MuJoCo for computation, and Isaac Sim for visualization.

Algorithm

Joint Position is $\mathbf{q} \in \mathbb{R}^{n_v}$ (where $n_v = 21$, representing the 21 DOF, specfically model.nv is the dimension of the joint velocity). Targets are world positions for the five fingertip bodies (targets dict). Only fingertip positions are constrained (jacp from mj_jacBody; jacr, the fingertip orientaton, is unused).

Iteration loop

Repeat up to max_iter times / return early if error lower than tol tolerance:

1. mj_forward(model, data) — compute forward kinematics with mujoco function

2. Build error $\mathbf{e}$ — for each fingertip that has a target, compute its error, err = target - pos; concatenate into one vector.

   For fingertip $i$:

   $$\mathbf{e}_i = \mathbf{p}_i^{\ast} - \mathbf{p}_i(\mathbf{q}) \in \mathbb{R}^3.$$

   Stacking $m$ fingertips:

   $$\mathbf{e} = \begin{bmatrix} \mathbf{e}_1 \\ \vdots \\ \mathbf{e}_m \end{bmatrix} \in \mathbb{R}^{3m}.$$

3. Convergence check — if $\max |\mathbf{e}| < \text{tol}$, exit early successfully.

4. Build Jacobian J — for each such fingertip, mj_jacBody → translation Jacobian jacp; vstack into J.

   Translation-only linearization for fingertip $i$:

   $$\mathrm{d}\mathbf{p}_i \approx \mathbf{J}_i\,\mathrm{d}\mathbf{q}, \qquad \mathbf{J}_i \in \mathbb{R}^{3 \times n_v}.$$

   Stacking:

   $$\mathbf{J} = \begin{bmatrix} \mathbf{J}_1 \\ \vdots \\ \mathbf{J}_m \end{bmatrix} \in \mathbb{R}^{3m \times n_v}.$$

5. Damped least squares — solve for a joint increment dq using damped least squares with $\lambda =$ damping.

   Compute $\Delta\mathbf{q}$ that minimizes (formalizing the optimization problem)

   $$\min_{\Delta\mathbf{q}} \; \bigl\|\mathbf{J}\,\Delta\mathbf{q} - \mathbf{e}\bigr\|_2^2 + \lambda\bigl\|\Delta\mathbf{q}\bigr\|_2^2.$$

   The minimizer satisfies the normal equations (The actual equation that we need to solve)

   $$\bigl(\mathbf{J}^{\mathsf T}\mathbf{J} + \lambda \mathbf{I}\bigr)\,\Delta\mathbf{q} = \mathbf{J}^{\mathsf T}\mathbf{e}.$$

6. Update — qpos[:nv] += step * dq (i.e. $\mathbf{q} \leftarrow \mathbf{q} + \alpha\,\Delta\mathbf{q}$, $\alpha =$ step).

7. Joint limits — if clip_limits, run clip_to_joint_limits, enforce joint limit

Return value: new joint pos, boolean on whether it successfully converge with error under tol, max error

Defaults: damping=1e-4, step=0.5, max_iter=200, tol=1e-4, clip_limits=True.

VR Teleoperation

Meta Quest hand and wrist tracking drives both arms of the bimanual rig in real time — a browser-side WebXR page streams hand poses over a secure WebSocket to a C++ ROS 2 bridge, which feeds a per-arm damped-least-squares IK controller in Isaac Sim, with a stereo POV feed rendered back to the headset so the operator sees through the robot's own cameras. The coolest part: it isn't sim-only. A gated flag streams the left arm's live IK solution straight to the physical robot over CAN, so the same pinch-to-grip, wrist-tracked teleop that drives the simulated arm can drive the real hardware in lockstep — headset in, robot arm moving, all over a USB tunnel with no Wi-Fi setup required.

Controls

Action	Effect
Move right / left wrist	Corresponding arm follows
Pinch thumb + index (either hand)	Corresponding gripper closes / opens
R (Isaac Sim window focused)	Recalibrate — re-home both arms to current wrist pose
--gain VALUE (default 1.0)	Scales arm motion per metre of real wrist motion

Reinforcement Learning

In-Hand Manipulation RL Task

Reinforcement learning-based in-hand cube reorientation task for the 16 DOF Wato hand in Isaac Lab. The policy learns to reorient a DexCube held in the palm toward commanded goal orientations using PPO (Proximal Policy Optimization).

Task Setup

Parameter	Value
Robot	16 DOF Wato hand (palm-up orientation)
Object	Isaac Nucleus DexCube, scale (0.8, 0.8, 0.8)
Goal command	Reorientation command, resampled on success
Success threshold	Orientation error < 0.4 rad
Episode length	20 seconds
Number of environments	2048 (default)
Simulation frequency	120 Hz
Action decimation	4

Reward Structure

The reward function combines task objectives with regularization penalties:

Category	Term	Weight	Description
Task Rewards	Position tracking	-3.0	L2 distance to goal position (hold in palm)
	Orientation tracking	10.0	Dense rotation signal: $1 / (\text{error} + 0.1)$
	Success bonus	50.0	Binary reward when error < 0.4 rad
	Object held bonus	0.5	+1/step when cube within 0.10 m of goal
	Angular velocity toward goal	0.2	Reward cube spin aligned with goal
	Spread activity	0.03	Encourages finger abduction (Wato hand)
Penalties	Object away penalty	-5.0	Terminal penalty when out of reach
	Action rate L2	-0.05	Penalize jerky commands
	Joint velocity L2	$-1 \times 10^{-4}$	Penalize high joint speeds
	Action L2	$-1 \times 10^{-4}$	Penalize large action magnitudes

Termination Conditions

Condition	Threshold
Time out	Episode length > 20 seconds
Max consecutive success	50 successful reorientations in one episode
Object out of reach	Cube drifts > 0.3 m from robot root
Orientation stagnation	Error stays > 0.5 rad for 150 consecutive steps

Training Details

• Algorithm: PPO with GAE (Generalized Advantage Estimation)
• Gamma: 0.998
• Entropy coefficient: 0.0001
• Steps per environment: 48
• Action smoothing: EMA joint-position targets, alpha = 0.85
• Max iterations: 5000

The task is implemented in Isaac Lab and adapted from their in-hand manipulation examples. Checkpoints are saved to logs/rsl_rl/wato_hand_cube/.

Additional RL Demonstrations

Locomotion RL

Reinforcement learning-based locomotion training.

Manipulation Task Setup

Parallelized manipulation task environments for RL training.

Geometric Fabrics PCA

Goal

Control a 21-DOF humanoid arm+hand using geometric fabrics — a control framework that's fast and stable without needing to solve an optimization problem every step.

The fingers have 15 joints, too many to command directly. So we use PCA to compress them to 7 numbers. PCA is trained on real human hand poses and learns the 7 directions that capture most of the variation in finger motion. You give it a 7-number command and it maps back to 15 joint angles via q_hand = W @ z + μ — just a change of coordinates. z is your 7D input, W maps it to joint space, and μ is the mean pose so that z = 0 gives you a natural resting hand.

Files

File	Purpose
compute_pca.py	Trains PCA from the HUST hand motion dataset, saves matrix/mean/bounds
humanoid_hand_pose_fabric.py	Core controller — builds all the fabric layers (arm, hand, collision, joint limits)
run_example.py	Runs the simulation loop at 60 Hz, renders optionally
humanoid_hand_params.yaml	All the tunable params — gains, collision spheres, joint limits
pca_matrix.npy, pca_mean.npy, pca_score_*.npy	Trained PCA files loaded at runtime
FABRICS/	NVIDIA geometric fabrics framework (submodule)

How It Works

Every step, four layers each compute forces and sum them together:

1. Arm attractor — holds the arm at a fixed default config (arm doesn't move)
2. Finger attractor — takes the 7D PCA command, maps to 15 finger joints, pulls toward them
3. Body repulsion — 21 spheres on the robot links push away from obstacles
4. Joint limit repulsion — pushes back when joints get near their limits

All of that sums into qdd = -M⁻¹ f, which gets integrated to get the next joint state.