PID Tuning

The AR4 simulation uses gravity-enabled Gazebo with gz_ros2_control, which requires carefully tuned PID gains. The upstream AR4 driver ships gains tuned for zero-gravity visualization (P=10, D=1, I=1) — unusable under real physics.

The Problem

The upstream URDF has no <dynamics> elements (damping=0, friction=0), and the implicitSpringDamper Gazebo tag has no effect with zero stiffness. Under gravity, every joint sags — joint 5 (wrist pitch) dropped 27° with the upstream gains.

Our sim_tabletop.launch.py post-processes the xacro output to inject <dynamics damping="..." friction="..."> into each joint via xml.etree.ElementTree, then spawns the URDF via file (not topic) to avoid DDS discovery issues.

Manual Tuning (Rounds 1-3)

Three rounds of manual tuning improved things but couldn't eliminate the J5 sag:

Automated Tuning with Optuna (Round 4)

We built an automated pipeline using Optuna TPE optimization, decoupled from ROS via Zenoh.

Architecture

The optimizer is a standalone Python container (no ROS dependency) that communicates with the simulation exclusively through Zenoh pub/sub. The tuning bridge node inside the sim container translates between Zenoh messages and ROS 2 controller operations.

Trial Sequence

Each trial (~65 seconds):

1. Sample 24 PID parameters (P, I, D, i_clamp × 6 joints) using log-uniform TPE
2. Publish gains to ar4/tuning/gains
3. Reconfigure — deactivate JTC, set new params via ros2 param set, reactivate
4. Home hold — command home pose, observe steady-state error for 15s
5. Reach move — command reach pose [0.5, 0.3, -0.4, 0.8, -0.5, 0.3] rad, observe for 25s
6. Score — cost = 10 × Σ(steady_state_errors) + 1 × settling_time

Search Space

24 total parameters (4 per joint × 6 joints).

Results (300 Trials)

Best trial: 162, cost = 40.726

All 300 trials completed successfully with zero penalties. The cost converged to a tight range of 40.72–40.76 after ~100 trials.

Optimized Gains

Key observations:

• J5 (wrist pitch) needs the highest P gain (6383) to counteract gravitational torque on the wrist
• J2 (shoulder) and J6 (wrist yaw) rely heavily on integral action for gravity compensation
• J1 (base) and J3 (elbow) need surprisingly low gains — they're well-supported by the arm structure
• The cost floor at ~40.7 reflects residual steady-state error that PID alone cannot eliminate (would need feedforward gravity compensation)

Comparison: Manual vs Optuna

Running the Tuning Pipeline

# Prerequisites
docker compose build overlay pid-tuner
 
# Run (default 300 trials)
./scripts/run_pid_tuning.sh
 
# Or specify trial count
./scripts/run_pid_tuning.sh 50
 
# Results
cat data/tuning/best_gains.yaml      # Best gains (YAML)
ls data/tuning/optuna_study.db        # Optuna study (SQLite, resumable)

The study is resumable — running the script again will continue from where it left off (same SQLite database).

Docker Services

Key Files

CycloneDDS Configuration

This host has multiple network interfaces (Docker bridges, Tailscale, Cloudflare WARP) which confuse CycloneDDS multicast discovery. Two separate configs handle this:

• Sim containers use docker/cyclonedds.xml (bundled at /etc/ros/cyclonedds.xml) — forces loopback interface
• Zenoh bridge uses zenoh/cyclonedds-bridge.xml — same loopback restriction but without the SharedMemory element (unsupported by the bridge's bundled CycloneDDS)