DynaTraj — Training and Data Collection (SB3 + dm_control)

This repo provides two entry-point scripts:

• sb3_train.py: Train an RL policy with Stable-Baselines3 on dm_control tasks and periodically save policy checkpoints
• sb3_collect.py: Load specific checkpoints and collect fixed-length trajectories into .npz datasets with minimal metadata

---

Environment & Dependencies

• Python 3.10+
• Required:
  • dm_control + MuJoCo: pip install dm-control mujoco
  • Reinforcement learning: pip install stable-baselines3
  • Utilities: pip install numpy tqdm torch
  • Gym API: prefers gymnasium, falls back to gym (handled automatically)
• Optional (only if you need on-screen rendering during collection):
  • OpenCV UI: pip install opencv-python
  • or Matplotlib: pip install matplotlib

Note: On headless Linux/remote environments, rendering might require EGL/OSMesa configuration. If you do not render, you do not need any display backend.

---

Output Layout (Conventions)

• Training checkpoints: weights/<domain>/<task>/ckpt-<k>.pt
  • k starts from 1; default save interval is every 10,000 timesteps
  • Examples: ckpt-1.pt ≈ 10k steps, ckpt-2.pt ≈ 20k steps
• Full SB3 model: weights/sb3_<algo>_<domain>-<task>_seed<seed>_<timestamp>.zip
• Collected datasets: dataset/*.npz with a companion *_metadata.pkl

Default absolute paths in this repo:

• Checkpoint root: /home/lau/sim/DynaTraj/weights
• Dataset output: /home/lau/sim/DynaTraj/dataset

---

Training

Script: sb3_train.py

Supported algorithms (argument is case-sensitive and must be lowercase here):

• sac, ppo, td3

Common args:

• --domain: dm_control domain, e.g., cheetah, quadruped (default: cheetah)
• --task: task name, e.g., run, walk (default: run)
• --algo: sac|ppo|td3 (default: sac)
• --total_timesteps: total training steps (default: 500000)
• --n_envs: number of parallel envs (default: 1; uses sub-process vectorization if >1)
• --seed: random seed (default: 0)
• --device: cpu|cuda|auto (default: auto)
• --out_dir: where to save models & checkpoints (default: /home/lau/sim/DynaTraj/weights)

Example:

python /home/lau/sim/DynaTraj/train_sb3_dmcontrol.py \
  --domain cheetah --task run\
  --algo sac \
  --total_timesteps 500000 \
  --out_dir /home/lau/sim/DynaTraj/weights

Checkpointing:

• A policy checkpoint is saved every 10,000 global timesteps as ckpt-<k>.pt (1-based counter).
• A full SB3 model .zip is also saved at the end for plain SB3 loading if needed.

---

Data Collection

Script: sb3_collect.py

Purpose: Run inference with specific training checkpoints and write fixed-length trajectories to .npz files plus metadata.

Common args:

• --domain / --task: must match training (defaults: cheetah / run)
• --ckpt_root: root directory of checkpoints (default: /home/lau/sim/DynaTraj/weights)
• --ckpt_indices: comma-separated checkpoint indices, e.g., 1,5,10 (note: starts from 1)
• --trajectories_per_ckpt: how many trajectories per checkpoint (default: 5120)
• --steps_per_trajectory: steps per trajectory (default: 24)
• --out_dir: dataset output directory (default: /home/lau/sim/DynaTraj/dataset)
• --device: inference device (default: cpu)
• --render: optional flag to render frames (requires OpenCV or Matplotlib)
• --algo: usually unnecessary. The script reads the real algo name from the checkpoint payload. Only used as a fallback (must be UPPERCASE: SAC|PPO|TD3).

Example:

python sb3_collect.py \
--domain cheetah --task run \
--algo SAC \
--ckpt_root ./weights \
--ckpt_indices 1,20,30,50 \
--trajectories_per_ckpt 1024 \
--steps_per_trajectory 512 \
--out_dir ./dataset \
--device cpu

With rendering (optional):

python /home/lau/sim/DynaTraj/sb3_collect.py \
--domain cheetah --task run \
--algo SAC \
--ckpt_root /home/lau/sim/DynaTraj/weights \
--ckpt_indices 50 \
--trajectories_per_ckpt 5120 \
--steps_per_trajectory 24 \
--out_dir /home/lau/sim/DynaTraj/dataset \
--device cpu \
  --render

For Bouncing ball: python bb_collect.py --trajectories 1024 --steps_per_trajectory 8192

Notes:

• The script searches checkpoints under ckpt_root/<domain>/<task>/ as ckpt-<k>.pt. If you see an error for ckpt-0.pt, switch to 1-based indices.
• If your environment has no display backend, simply omit --render.

---

Output Format

For each k:

• Dataset file: <out_dir>/sb3_<domain>_<task>_ckpt<kNNN>_<timestamp>.npz
• Metadata file: <out_dir>/sb3_<domain>_<task>_ckpt<kNNN>_<timestamp>_metadata.pkl

Metadata keys include:

• domain, task, algo, seed
• ckpt_index (the k you collected)
• trajectories_per_ckpt, steps_per_trajectory
• total_trajectories, total_steps
• render (whether rendering was enabled)

The exact arrays inside .npz are defined by the internal TrajectoryBuffer implementation. Typically they include time-aligned observation/state tensors, actions, rewards, and done flags.

---

Data Format

Transition:

a_t = pi(s_t)
s_t+1 = f(s_t, a_t)

1.Cheetah

• state
  • qpos_t(x,z,pitch,joint) = 9
  • qvel_t(x,z,pitch,joint) = 9
• action
  • tau_t(joint) = 6

Tips

• Prefer absolute paths (all examples above use absolute paths).
• Training uses DummyVecEnv/SubprocVecEnv, automatically flattens dm_control observations and clips actions to env bounds.
• Collection reconstructs an SB3 policy and loads the policy_state_dict from each checkpoint; it does not require the full .zip model.
• Algo-arg case: training expects lowercase (sac|ppo|td3), collection fallback uses UPPERCASE (SAC|PPO|TD3).