Figure-8 Robot Tracking — Setup Guide

This project uses MuJoCo, Gymnasium, and Stable-Baselines3 for reinforcement learning–based robotic control.

⚠️ This setup was tested on Ubuntu 22.04 running inside WSL2 (Windows Subsystem for Linux).
It should also work on native Ubuntu 22.04 / 24.04.

---

1. System Dependencies (Ubuntu)

Update package lists and install required system libraries:

sudo apt-get update
sudo apt-get install -y git python3 python3-venv python3-pip
sudo apt-get install -y libglfw3 libglew-dev libgl1-mesa-glx libosmesa6

2. Create Python Virtual Environment

Create and activate a clean Python virtual environment:

python3 -m venv .venv
source .venv/bin/activate
python -m pip install -U pip

3. Install Python Dependencies

Install MuJoCo, Gymnasium, and Stable-Baselines3:

pip install mujoco gymnasium "stable-baselines3[extra]"

4. Robot Assets (MuJoCo Menagerie)

The Franka Emika Panda model is sourced from the official DeepMind Menagerie:

git clone https://github.com/google-deepmind/mujoco_menagerie.git

5. Automated Setup (Optional)

If a setup script is provided, run:

bash setup.sh

🏋️ Training

To train the PPO agent:

python train_ppo.py

👁️ Evaluation / Viewer

To run a trained model and visualize the rollout:

python eval_rollout_viewer.py

⚠️ Usage Reminder

Always activate the virtual environment before running training or rollout scripts:

source .venv/bin/activate

📊 Performance Validation

To evaluate the policy, we conducted inference tests across different spatial scales and temporal frequencies.

Important

Visualization Key:

• Red Line: Target trajectory.
• Green Line: Actual end-effector path.

---

▶ Click to view Inference Testing: Small Out-of-Distribution (1.0 Hz vs 1.3 Hz)

1.0 Hz Frequency	1.3 Hz Frequency
Mean Position Error: 2.17 cm	Mean Position Error: 2.11 cm

Note on Methodology: Results demonstrate zero-shot generalization to a compressed figure-8 scale ($0.10 \times 0.06$ m) entirely unseen during the training phase. To ensure statistical significance, reported metrics are the grand mean derived from 10 randomized episodes. For each 20-second test rollout, a temporal average of the tracking error is computed, followed by an ensemble average across all 10 episodes.

▶ Click to view Inference Testing: Large Out-of-Distribution (1.0 Hz vs 1.3 Hz)

1.0 Hz Frequency	1.3 Hz Frequency
Mean Position Error: 4.99 cm	Mean Position Error: 6.80 cm

Note on Methodology: Results demonstrate zero-shot generalization to an expanded figure-8 scale ($0.22 \times 0.15$ m). Reported metrics are the grand mean of 10 randomized 20-second episodes.

Kinematic Limit Analysis: Our testing identifies a critical physical boundary for the Franka Panda. For any trajectory where the scale parameter $a > 0.20$ m, the robot arm reaches the edge of its operational workspace. At these scales, the joints (specifically Joints 4 and 5) encounter hardware saturation, preventing the end-effector from completing the full path. The increased error and 0% success rate at 1.3 Hz reflect this mechanical constraint rather than policy divergence.

▶ Click to view Inference Testing: In-Distribution Baseline (1.0 Hz vs 1.3 Hz)

1.0 Hz Frequency (Baseline)	1.3 Hz Frequency (Stress Test)
Mean Position Error: 2.62 cm	Mean Position Error: 2.86 cm

Note on Methodology: These results represent the In-Distribution performance on the training scale ($0.14 \times 0.10$ m). Reported metrics are the grand mean derived from 10 randomized episodes. For each 20-second test rollout, a temporal average of the tracking error is computed, followed by an ensemble average across all 10 episodes to ensure statistical robustness.

▶ Click to view Inference Testing: Specialized Policy (0.75 Hz Baseline)

In-Distribution Testing (0.75 Hz)
Mean Position Error: 1.84 cm

Note on Methodology: These results represent the Specialized Policy performance on the standard training scale ($0.14 \times 0.10$ m) at the target frequency of 0.75 Hz. To ensure statistical significance, reported metrics are the grand mean derived from 50 randomized episodes. For each 20-second test rollout, a temporal average of the tracking error is computed, followed by an ensemble average across all 50 episodes.