Training & Evaluation¶

Train on flat terrain¶

This is the recommended starting point.

cd ~/rlgpu_ws/IsaacLab
./isaaclab.sh -p ~/Desktop/guguji_simulation/guguji_isaaclab/scripts/rsl_rl/train.py \
  --task=Isaac-Velocity-Flat-Guguji-v0 \
  --num_envs=4096 \
  --headless

The flat task uses a forward-velocity curriculum that begins at 0.10 m/s and scales toward 0.30 m/s as tracking improves.

Train on rough terrain¶

After the flat policy becomes stable, move to the rough-terrain task.

cd ~/rlgpu_ws/IsaacLab
./isaaclab.sh -p ~/Desktop/guguji_simulation/guguji_isaaclab/scripts/rsl_rl/train.py \
  --task=Isaac-Velocity-Rough-Guguji-v0 \
  --num_envs=2048 \
  --headless

Evaluate a trained policy¶

cd ~/rlgpu_ws/IsaacLab
./isaaclab.sh -p ~/Desktop/guguji_simulation/guguji_isaaclab/scripts/rsl_rl/play.py \
  --task=Isaac-Velocity-Flat-Guguji-Play-v0 \
  --num_envs=50

Load a specific checkpoint¶

cd ~/rlgpu_ws/IsaacLab
./isaaclab.sh -p ~/Desktop/guguji_simulation/guguji_isaaclab/scripts/rsl_rl/play.py \
  --task=Isaac-Velocity-Flat-Guguji-Play-v0 \
  --num_envs=50 \
  --load_run=2026-04-17_15-33-29 \
  --checkpoint=model_19999.pt

Logging and artifacts¶

Training logs and checkpoints are written to:

logs/rsl_rl/<experiment_name>/<timestamp>/

The evaluation script also exports:

• policy.pt for TorchScript deployment
• policy.onnx for ONNX-based downstream use

Inspect training with TensorBoard¶

cd ~/rlgpu_ws/IsaacLab
./isaaclab.sh -p -m tensorboard.main --logdir=logs/rsl_rl/guguji_flat

cd ~/rlgpu_ws/IsaacLab
./isaaclab.sh -p -m tensorboard.main --logdir=logs/rsl_rl/guguji_rough

Practical workflow¶

1. Verify environment registration before starting long runs.
2. Train the flat policy first.
3. Check whether the robot walks straight instead of circling.
4. Export the best checkpoint and validate it with play.py.
5. Then move to rough terrain and iterate on reward shaping only if needed.

Common training focus areas¶

1. Straight walking¶

The project explicitly penalizes yaw rate and lateral drift to avoid policies that appear to move forward in body frame while actually circling.

2. Gait symmetry¶

Recent reward shaping emphasizes left-right alternation and knee symmetry so the policy does not collapse into a limping pattern.

3. Higher-quality swing phase¶

Reference gait amplitudes and swing knee scaling were increased to encourage clearer leg lift and more visible step alternation.