JackBot ML training and evaluation

Quick-start

1. Create a Python virtualenv and activate it:

```bash
python -m venv .venv
source .venv/bin/activate
```

2. Install dependencies (GPU users should install `torch` appropriate for their CUDA):

```bash
pip install -r requirements.txt
```

3. Quick training (short, for smoke test):

```bash
python ml/run_train.py --timesteps 50000 --model ml/checkpoints/ppo_joint_command
```

Use multi-robot training with `--num-robots` and `--robot-spacing`:

```bash
python ml/run_train.py --timesteps 50000 --model ml/checkpoints/ppo_joint_command --num-robots 3 --robot-spacing 0.75
```

Choose the start pose at reset with `--start-pose`:

```bash
python ml/run_train.py --timesteps 50000 --model ml/checkpoints/ppo_joint_command --start-pose init_deg
```

If you want to watch the agent train in the PyBullet window, add `--gui`:

```bash
python ml/run_train.py --timesteps 50000 --model ml/checkpoints/ppo_joint_command --gui
```

4. Evaluate the trained model in GUI mode:

```bash
python ml/run_eval.py --model ml/checkpoints/ppo_joint_command.zip --episodes 3 --gui
```

For evaluation with multiple robots and start pose:

```bash
python ml/run_eval.py --model ml/checkpoints/ppo_joint_command.zip --episodes 3 --gui --num-robots 2 --robot-spacing 0.6 --start-pose init_deg
```

Notes

- `ml/env.py` exposes a `JackBotEnv` gym environment that uses your `JackBotUrdf.urdf`.
- The observation is `[joint_angles..., vx, vy, vz, omega]` and the action is per-joint delta in [-1,1].
- Start with small timesteps and in `use_gui=False` to speed up iteration. Increase timesteps and tune the reward for better walking quality.
- Keep the trained models off hardware until they behave well in simulation.