fix: merge conflict

.devcontainer/devcontainer.json

@@ -34,8 +34,7 @@
     },
     "remoteUser": "vscode",
     "runArgs": [
-        "--gpus",
+        "--device", "nvidia.com/gpu=all"
-        "all"
     ],
     // Ensure the .venv persists using a named volume for performance and parity
     "mounts": [

.github/workflows/update_hpc_requirements.yml

@@ -0,0 +1,38 @@
+name: Update HPC requirements
+
+on:
+  push:
+    paths:
+      - pyproject.toml
+    branches:
+      - main
+      - dev
+      - "ci/**"
+
+jobs:
+  update-hpc-requirements:
+    runs-on: ubuntu-latest
+    permissions:
+      contents: write
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          ref: ${{ github.head_ref || github.ref_name }}
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+
+      - name: Install uv
+        uses: astral-sh/setup-uv@v5
+
+      - name: Regenerate env/hpc/requirements.txt
+        run: uv run scripts/export_hpc_requirements.py
+
+      - name: Commit updated requirements if changed
+        uses: stefanzweifel/git-auto-commit-action@v5
+        with:
+          commit_message: "chore(hpc): update env/hpc/requirements.txt from pyproject.toml [skip ci]"
+          file_pattern: env/hpc/requirements.txt
+          commit_author: "github-actions[bot] <github-actions[bot]@users.noreply.github.com>"

.gitignore

@@ -1,6 +1,7 @@
 # Model files
 artifacts/*
 runs/*
+wandb/
 
 # Python-generated files
 __pycache__/
@@ -372,7 +373,6 @@ celerybeat.pid
 # Environments
 .env
 .venv
-env/
 venv/
 ENV/
 env.bak/
@@ -472,7 +472,6 @@ tags
 [Ll]ib
 [Ll]ib64
 [Ll]ocal
-[Ss]cripts
 pyvenv.cfg
 .venv
 pip-selfcheck.json
@@ -519,3 +518,4 @@ Icon
 Network Trash Folder
 Temporary Items
 .apdisk
+*.pdf

.vscode/settings.json

@@ -0,0 +1,7 @@
+{
+    "python.testing.pytestArgs": [
+        "tests"
+    ],
+    "python.testing.unittestEnabled": false,
+    "python.testing.pytestEnabled": true
+}

README.md

@@ -15,3 +15,7 @@ example command:
 ```bash
 uv run src/train.py --model_name my_model --epochs 50 --batch_size 32
 ```
+
+## HPC
+
+See **[docs/HPC.md](docs/HPC.md)** for the full guide, including environment setup, cluster selection, interactive debugging, and job submission.

configs/hpc/smoke_test.yaml

@@ -0,0 +1,11 @@
+# Minimal config to verify HPC setup is functional.
+# Run with: python src/train.py --config-path configs/hpc/smoke_test.yaml
+exp_name: "hpc_smoke_test"
+seed: 0
+track: false           # Test WandB integration
+capture_video: false  # No rendering for smoke test
+save_model: true      # Test the end-of-training save routine
+num_envs: 512
+total_timesteps: 65536
+num_steps: 128
+cuda: true

configs/production_training.yaml

@@ -0,0 +1,24 @@
+# Full PPO training config for Brittle Star (HPC Production)
+exp_name: "production_training"
+seed: 1
+track: true
+capture_video: true
+save_model: true
+checkpoint_frequency: 100  # not yet implemented in train.py but here for future use
+
+# Scaling for HPC (using A100 GPU slices)
+num_envs: 128
+total_timesteps: 10000000
+num_steps: 128
+num_minibatches: 4
+update_epochs: 4
+
+# Algorithm
+learning_rate: 2.5e-4
+anneal_lr: true
+gamma: 0.99
+gae_lambda: 0.95
+clip_coef: 0.1
+ent_coef: 0.01
+vf_coef: 0.5
+cuda: true

docs/CONTRIBUTING.md

@@ -32,4 +32,4 @@ Code readability is paramount, as code is read far more frequently than it is wr
 * **Algorithms & Frameworks:** Proximal Policy Optimization (PPO) is the recommended baseline algorithm. CleanRL should be used as a starting point and adapted for continuous action spaces. All Artificial Neural Network (ANN) controller architectures must be implemented using Flax.
 * **Simulation:** The simulation environment utilizes a MuJoCo brittle star. XML MuJoCo structures must remain realistic and respect morphological constraints.
 * **Experiment Tracking:** Weights & Biases (wandb) must be utilized for tracking and logging all experiments.
-* **Code Styling:** All code must conform to the chosen style guide (i.e. Google standard). This is enforced using build tools and pre-commit hooks such as flake8, black, or isort.
+* **Code Styling:** All code must conform to the chosen style guide (Google standard). This is enforced via `uv` using **ruff** and pre-commit hooks.

docs/HPC.md

@@ -0,0 +1,85 @@
+# HPC Guide
+
+Full documentation: <https://docs.hpc.ugent.be/>
+
+## Storage Overview
+
+- **Run Outputs**: Written to `$VSC_SCRATCH` during the job (fast I/O) and copied to `$VSC_DATA` at the end for persistence.
+- **Virtual Environments**: Managed on **`$VSC_DATA`** by mirroring configuration files. This avoids the 3GB home quota without requiring symlinks in the project root.
+
+## Initial Environment Setup
+
+Run **once** after cloning the repository. This script handles all modules, mirroring, and environment synchronization.
+
+```bash
+# Option A: Interactive (on a compute node)
+module swap cluster/donphan  # Debug cluster (CPU only)
+# OR for GPU clusters:
+# module swap cluster/joltik
+# module swap cluster/accelgor
+# module swap cluster/litleo
+
+qsub -I -l nodes=1:gpus=1  # Only for GPU clusters
+cd "${PBS_O_WORKDIR}"
+bash scripts/hpc/install.sh
+
+# Option B: Batch (Run in background)
+# NOTE: GPU clusters (joltik/accelgor/litleo) require -l gpus=1 at runtime
+qsub -l gpus=1 scripts/hpc/install.sh
+```
+
+## Production vs. Debug Clusters
+
+Our scripts are cluster-agnostic and do **not** have hardcoded GPU requirements. Instead, you must request GPUs at runtime using the `-l gpus=1` flag when submitting to a production GPU cluster.
+
+### Debugging (Donphan)
+The `donphan` cluster does not support GPUs. Simply run the scripts without extra resource flags:
+```bash
+module swap cluster/donphan
+qsub scripts/hpc/train.pbs
+```
+
+### Production (Joltik, Accelgor, Litleo)
+These clusters provide GPU acceleration and **require** a GPU request at runtime:
+```bash
+module swap cluster/joltik  # or accelgor/litleo
+qsub -l gpus=1 scripts/hpc/train.pbs
+```
+
+## Interactive Debugging
+
+To activate your environment for interactive work, simply run the same `install.sh` script.
+
+```bash
+qsub -I -l nodes=1:ppn=4 -l walltime=1:00:00
+cd "$PBS_O_WORKDIR"
+bash scripts/hpc/install.sh
+```
+
+### Verification Commands
+
+After installation, run these commands to ensure your environment is set up correctly:
+
+1. **Verify Quota Safety**:
+   ```bash
+   ls -d venvs 2>/dev/null && echo "FAIL" || echo ">>> PASS: Project root is clean."
+   ```
+2. **Verify Library Versions (NumPy Fix)**:
+   ```bash
+   python -c "import numpy; print(f'NumPy: {numpy.__version__}')"
+   # Expected: 2.x.x (Venv version), not 1.2x (System version)
+   ```
+3. **Verify GPU Access**:
+   ```bash
+   python -c "import torch, jax; print(f'GPU: {torch.cuda.is_available()}'); print(f'JAX: {jax.devices()}')"
+   ```
+
+## Managing Dependencies
+
+`env/hpc/requirements.txt` is auto-generated from `pyproject.toml`. To regenerate:
+
+```bash
+uv run scripts/export_hpc_requirements.py
+```
+
+Modules listed in `env/hpc/modules.txt` are automatically excluded from the pip requirements to save space and use HPC-optimized binaries.

env/hpc/modules.txt

@@ -0,0 +1,3 @@
+GCCcore/13.3.0
+Python/3.12.3-GCCcore-13.3.0
+FFmpeg/7.0.2-GCCcore-13.3.0

env/hpc/requirements.txt

@@ -0,0 +1,19 @@
+biorobot==0.4.2
+cleanrl>=0.4.8
+evosax==0.2.0
+flax>=0.12.2
+gymnasium>=1.2.3
+ipykernel==7.2.0
+jax[cuda13]==0.9.0.1
+numpy>=2.0.0
+protobuf>=5.0.0
+warp-lang
+mujoco-warp
+matplotlib==3.10.8
+mediapy==1.2.6
+optax>=0.2.6
+pyopengl>=3.1.10
+pyopengl-accelerate>=3.1.10
+tyro>=1.0.10
+wandb==0.24.2
+torch>=2.4.0

experiments/_train_backup.py

@@ -1,4 +1,8 @@
+import datetime
 import random
+import yaml
+import subprocess
+import sys
 import time
 from dataclasses import asdict
 from functools import partial
@@ -7,6 +11,7 @@ from typing import Callable
 import flax
 import jax
 import jax.numpy as jnp
+import matplotlib.pyplot as plt
 import numpy as np
 import optax
 import torch
@@ -35,11 +40,11 @@ def convert_obs_dict_to_array(obs_dict: dict) -> jnp.ndarray:
     )
 
 
-def make_env(config_path: str | None, num_envs: int) -> Callable:
+def make_env(env_config_path: str | None, num_envs: int) -> Callable:
     def thunk():
-        if config_path is None:
+        if env_config_path is None:
             return BrittleStarJaxEnvWrapper.default(num_envs=num_envs)
-        return BrittleStarJaxEnvWrapper.from_config(config_path, num_envs=num_envs)
+        return BrittleStarJaxEnvWrapper.from_config(env_config_path, num_envs=num_envs)
 
     return thunk
 
@@ -63,11 +68,29 @@ def save_model(model_path: str, agent_state: TrainState, args: PPOArgs):
 def train(args: PPOArgs):
     args.batch_size = args.num_envs * args.num_steps
     args.minibatch_size = args.batch_size // args.num_minibatches
+    args.num_iterations = args.total_timesteps // args.batch_size
+
+    # Try to get git short hash
+    try:
+        git_hash = (
+            subprocess.check_output(["git", "rev-parse", "--short", "HEAD"]).decode("ascii").strip()
+        )
+    except Exception:
+        git_hash = "none"
+
+    run_name = f"{args.exp_name}__seed_{args.seed}__{git_hash}__{int(time.time())}"
     # args.num_iterations = args.total_timesteps // args.batch_size
     args.num_iterations = 5
     run_name = f"{args.exp_name}__seed_{args.seed}__{int(time.time())}"
     print(f"running name: {run_name}")
 
+    if args.run_dir is None:
+        args.run_dir = f"runs/{run_name}"
+
+    import os
+
+    os.makedirs(args.run_dir, exist_ok=True)
+
     if args.track:
         import wandb
 
@@ -80,7 +103,7 @@ def train(args: PPOArgs):
             save_code=True,
         )
 
-    writer = SummaryWriter(f"runs/{run_name}")
+    writer = SummaryWriter(args.run_dir)
     writer.add_text(
         "hyperparameters",
         "|param|value|\n|---|---|\n" + "\n".join(f"|{k}|{v}|" for k, v in vars(args).items()),
@@ -96,7 +119,7 @@ def train(args: PPOArgs):
     print(f"Running on device: {device}")
 
     print("Creating the environment...")
-    env = make_env(config_path=args.config_path, num_envs=args.num_envs)()
+    env = make_env(env_config_path=args.env_config_path, num_envs=args.num_envs)()
     print(f"Environment: {env}")
 
     episode_stats = EpisodeStatistics(
@@ -235,9 +258,13 @@ def train(args: PPOArgs):
 
     # Reset once to get initial state
     print("Resetting the environment...")
+    if not sys.stdout.isatty():
+        print(f">>> [HPC] Initial reset started: {time.ctime()}", flush=True)
     next_env_state = env.reset(seed=args.seed)
     next_obs = convert_obs_dict_to_array(next_env_state.observations)
     next_done = jnp.zeros(args.num_envs, dtype=jnp.bool_)
+    if not sys.stdout.isatty():
+        print(f">>> [HPC] Initial reset completed: {time.ctime()}", flush=True)
 
     def step_once(carry, _, env_step_fn):
         agent_state, episode_stats, obs, done, key, env_state = carry
@@ -277,28 +304,45 @@ def train(args: PPOArgs):
     )
 
     print("Starting training...")
-    iters_bar = tqdm.tqdm(range(1, args.num_iterations + 1))
+    iters_bar = tqdm.tqdm(
+        range(1, args.num_iterations + 1),
+        disable=not sys.stdout.isatty(),
+    )
     returns = []
-    for _ in iters_bar:
+    is_tty = sys.stdout.isatty()
+    for iteration in iters_bar:
         iteration_time_start = time.time()
 
+        if not is_tty and iteration == 1:
+            print(f">>> [HPC] Starting first rollout (JIT): {time.ctime()}", flush=True)
+
         agent_state, episode_stats, next_obs, next_done, storage, key, next_env_state = rollout(
             agent_state, episode_stats, next_obs, next_done, key, next_env_state
         )
 
+        if not is_tty and iteration == 1:
+            print(f">>> [HPC] First rollout completed: {time.ctime()}", flush=True)
+
         global_step += args.num_steps * args.num_envs
         storage = compute_gae(agent_state, next_obs, next_done, storage)
+
+        if not is_tty and iteration == 1:
+            print(f">>> [HPC] Starting first PPO update (JIT): {time.ctime()}", flush=True)
+
         agent_state, loss, pg_loss, v_loss, entropy_loss, approx_kl, key = ppo_instance.update_ppo(
             agent_state, storage, key
         )
 
+        if not is_tty and iteration == 1:
+            print(f">>> [HPC] First PPO update completed: {time.ctime()}", flush=True)
+
+        losses.append(jnp.mean(loss))
+
         avg_episodic_return = np.mean(jax.device_get(episode_stats.returned_episode_returns))
         iters_bar.set_postfix_str(
             f"global_step={global_step}, avg_episodic_return={avg_episodic_return}"
         )
 
-        returns.append(avg_episodic_return)
-
         writer.add_scalar("charts/avg_episodic_return", avg_episodic_return, global_step)
         writer.add_scalar(
             "charts/avg_episodic_length",
@@ -325,9 +369,39 @@ def train(args: PPOArgs):
             global_step,
         )
 
+        if not is_tty:
+            sps = int(global_step / (time.time() - start_time))
+            remaining_steps = args.total_timesteps - global_step
+            eta_seconds = int(remaining_steps / sps) if sps > 0 else 0
+            eta_str = str(datetime.timedelta(seconds=eta_seconds))
+
+            print(
+                f"Iteration {iteration}/{args.num_iterations} | "
+                f"Step {global_step}/{args.total_timesteps} | "
+                f"SPS {sps} | "
+                f"Return {avg_episodic_return:.4f} | "
+                f"ETA {eta_str}",
+                flush=True,
+            )
+
     if args.save_model:
         model_path = f"runs/{run_name}/{args.exp_name}.cleanrl_model"
         save_model(model_path, agent_state, args)
+        model_path = f"{args.run_dir}/{args.exp_name}.cleanrl_model"
+        with open(model_path, "wb") as f:
+            f.write(
+                flax.serialization.to_bytes(
+                    [
+                        vars(args),
+                        [
+                            agent_state.params["sensor_params"],
+                            agent_state.params["actor_params"],
+                            agent_state.params["critic_params"],
+                            agent_state.params["feature_extractor_params"],
+                        ],
+                    ]
+                )
+            )
         print(f"model saved to {model_path}")
 
     env.close()
@@ -340,10 +414,29 @@ def train(args: PPOArgs):
         show_window=True,
         filename=f"runs/{run_name}/{args.exp_name}_losses.png",
     )
+    plt.plot(losses)
+    plt.title("PPO Loss, mean over minibatches")
+    plt.savefig(f"{args.run_dir}/{args.exp_name}_losses.png")
+    plt.close()
 
 
 def main() -> None:
-    args = tyro.cli(PPOArgs)
+    temp_args = tyro.cli(PPOArgs)
+
+    if temp_args.env_config_path is not None:
+        with open(temp_args.env_config_path, "r") as f:
+            config = yaml.safe_load(f)
+            if config:
+                # parse PPOArgs with defaults from yaml.
+                for key, value in config.items():
+                    if hasattr(temp_args, key):
+                        setattr(temp_args, key, value)
+
+                # Re-parse CLI to ensure they OVERRIDE the yaml
+                args = tyro.cli(PPOArgs, default=temp_args)
+    else:
+        args = temp_args
+
     train(args)
 
 

pyproject.toml

@@ -12,6 +12,10 @@ dependencies = [
     "gymnasium>=1.2.3",
     "ipykernel==7.2.0",
     "jax==0.9.0.1",
+    "numpy>=2.0.0",
+    "protobuf>=5.0.0",
+    "warp-lang",
+    "mujoco-warp",
     "matplotlib==3.10.8",
     "mediapy==1.2.6",
     "optax>=0.2.6",
@@ -19,12 +23,16 @@ dependencies = [
     "pyopengl-accelerate>=3.1.10",
     "tyro>=1.0.10",
     "wandb==0.24.2",
+    "torch>=2.4.0",
 ]
 
 [project.optional-dependencies]
 cuda = [
     "jax[cuda13]==0.9.0.1",
 ]
+analysis = [
+    "tensorboard",
+]
 
 [dependency-groups]
 dev = [

ruff.toml

@@ -1,4 +1,5 @@
 line-length = 100
+exclude = ["wandb"]
 
 [lint]
 extend-select = [

scripts/analysis/README.md

@@ -0,0 +1,27 @@
+# Experiment Analysis Tools
+
+This directory contains scripts for post-processing and analyzing experiment results, including TensorBoard logs and saved model weights.
+
+## Scripts
+
+### 1. `explore_tensorboard.py`
+A CLI tool to summarize TensorBoard `tfevents` files without a GUI.
+
+**Key Features:**
+- Displays last values, min, max, and step counts for all scalar metrics.
+- Calculates total run duration and estimated completion percentage.
+- Exports granular scalar data to CSV for analysis in Excel/Pandas.
+
+**Usage:**
+```bash
+# General usage
+python explore_tensorboard.py <run_directory>
+
+# Exporting data
+python explore_tensorboard.py <run_directory> --csv data.csv
+```
+
+**Requirements:**
+- `pandas`
+- `tensorboard`
+- `tensorflow-cpu` (or `tensorflow`)

scripts/analysis/explore_tensorboard.py

@@ -0,0 +1,143 @@
+#!/usr/bin/env python3
+"""
+Reproducible CLI tool to explore TensorBoard logs.
+Designed for both local development and HPC diagnostics.
+
+Requirements:
+    pip install tensorboard
+
+Usage:
+    python explore_tensorboard.py <path_to_run_directory> [--csv output.csv]
+"""
+
+import argparse
+import os
+import sys
+import csv
+
+try:
+    from tensorboard.backend.event_processing import event_accumulator
+except ImportError:
+    print("Error: Missing dependency. Please run: pip install tensorboard")
+    sys.exit(1)
+
+
+def explore_run(log_dir):
+    """
+    Extracts and displays a summary of scalar metrics from a TensorBoard log directory.
+    """
+    print(f"\n{'=' * 20} Exploring Run {'=' * 20}")
+    print(f"Directory: {log_dir}")
+    print(f"{'=' * 55}\n")
+
+    if not os.path.exists(log_dir):
+        print(f"Error: Directory '{log_dir}' does not exist.")
+        return None
+
+    # Initialize EventAccumulator
+    # size_guidance=0 loads all data points for each tag.
+    ea = event_accumulator.EventAccumulator(
+        log_dir,
+        size_guidance={
+            event_accumulator.SCALARS: 0,
+            event_accumulator.TENSORS: 0,
+        },
+    )
+
+    print("Loading event files (this may take a moment for large runs)...")
+    ea.Reload()
+
+    tags = ea.Tags()
+    scalar_tags = tags.get("scalars", [])
+
+    if not scalar_tags:
+        print("No scalar metrics found in this directory.")
+        return None
+
+    print(f"Found {len(scalar_tags)} scalar metrics.\n")
+
+    data = {}
+    summary = []
+
+    # Process scalar values
+    for tag in scalar_tags:
+        events = ea.Scalars(tag)
+        if not events:
+            continue
+
+        values = [e.value for e in events]
+        last_event = events[-1]
+        data[tag] = values
+
+        summary.append(
+            {
+                "Metric": tag,
+                "Steps": len(events),
+                "Last Value": f"{last_event.value:.4f}",
+                "Max": f"{max(values):.4f}",
+                "Min": f"{min(values):.4f}",
+            }
+        )
+
+    # Display summary table formatted manually
+    summary = sorted(summary, key=lambda x: x["Metric"])
+    print(f"{'Metric':<30} {'Steps':>10} {'Last':>12} {'Max':>12} {'Min':>12}")
+    print("-" * 80)
+    for row in summary:
+        print(
+            f"{row['Metric']:<30} {row['Steps']:>10} {row['Last Value']:>12} "
+            f"{row['Max']:>12} {row['Min']:>12}"
+        )
+
+    # Calculate and display global metadata
+    if "charts/SPS" in data:
+        sps_events = ea.Scalars("charts/SPS")
+        if len(sps_events) > 1:
+            total_duration_hours = (sps_events[-1].wall_time - sps_events[0].wall_time) / 3600
+            print(f"\nTotal Recorded Duration: {total_duration_hours:.2f} hours")
+
+            # Estimate completion if total_timesteps is available in hyperparameters
+            try:
+                hp_tags = [t for t in tags.get("tensors", []) if "hyperparameters" in t]
+                if hp_tags:
+                    hp_event = ea.Tensors(hp_tags[0])[0]
+                    hp_text = hp_event.tensor_proto.string_val[0].decode("utf-8")
+                    if "total_timesteps" in hp_text:
+                        for line in hp_text.split("\n"):
+                            if "total_timesteps" in line:
+                                target = int(line.split("|")[2].strip())
+                                current = ea.Scalars(scalar_tags[0])[-1].step
+                                percent = (current / target) * 100
+                                print(f"Progress: {current:,} / {target:,} steps ({percent:.1f}%)")
+            except Exception:
+                pass
+
+    return data
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Reproducible TensorBoard exploration tool.")
+    parser.add_argument("log_dir", help="Path to the TensorBoard run directory.")
+    parser.add_argument("--csv", help="Optional: Path to export scalar data to CSV.", default=None)
+
+    args = parser.parse_args()
+
+    scalar_data = explore_run(args.log_dir)
+
+    if args.csv and scalar_data:
+        # Reloading for wall_time and steps
+        ea = event_accumulator.EventAccumulator(args.log_dir).Reload()
+        with open(args.csv, mode="w", newline="") as f:
+            writer = csv.DictWriter(f, fieldnames=["tag", "step", "value", "wall_time"])
+            writer.writeheader()
+            for tag in scalar_data.keys():
+                for e in ea.Scalars(tag):
+                    writer.writerow(
+                        {"tag": tag, "step": e.step, "value": e.value, "wall_time": e.wall_time}
+                    )
+
+        print(f"\nData exported to: {args.csv}")
+
+
+if __name__ == "__main__":
+    main()

scripts/export_hpc_requirements.py

@@ -0,0 +1,86 @@
+#!/usr/bin/env python3
+"""Export HPC pip requirements from pyproject.toml.
+
+This is a LOCAL DEVELOPER UTILITY — run it on your own machine before pushing
+code whenever pyproject.toml dependencies change. It reads the modules from
+env/hpc/modules.txt and the full dependency list from pyproject.toml, then
+writes the remainder to env/hpc/requirements.txt.
+
+Usage:
+    uv run scripts/export_hpc_requirements.py
+"""
+
+from __future__ import annotations
+
+import re
+import sys
+from pathlib import Path
+
+ROOT = Path(__file__).parent.parent
+
+
+def normalise(name: str) -> str:
+    """Normalise a PyPI package name for comparison."""
+    return re.sub(r"[-_.]+", "-", name).lower()
+
+
+def pkg_name(dep: str) -> str:
+    """Extract the bare package name from a PEP 508 dependency string."""
+    return re.split(r"[\[=><~!;]", dep)[0].strip()
+
+
+def main() -> None:
+    import tomllib
+
+    modules_path = ROOT / "env" / "hpc" / "modules.txt"
+    if not modules_path.exists():
+        print(f"Error: {modules_path} not found.", file=sys.stderr)
+        sys.exit(1)
+
+    # Read normalized module names from base modules only
+    # Library modules (like PyTorch) are kept in requirements for portability
+    module_names = [
+        normalise(line.split()[0].split("/")[0])
+        for line in modules_path.read_text().splitlines()
+        if line.strip() and not line.startswith("#")
+    ]
+
+    pyproject_path = ROOT / "pyproject.toml"
+    with pyproject_path.open("rb") as f:
+        data = tomllib.load(f)
+
+    # Collect all dependencies, merging 'cuda' extras into base dependencies
+    dep_dict: dict[str, str] = {}
+    for dep in data.get("project", {}).get("dependencies", []):
+        dep_dict[normalise(pkg_name(dep))] = dep
+
+    # Add cuda extras (takes precedence for HPC)
+    optional_deps = data.get("project", {}).get("optional-dependencies", {})
+    for group in ["cuda"]:
+        for dep in optional_deps.get(group, []):
+            dep_dict[normalise(pkg_name(dep))] = dep
+
+    deps = list(dep_dict.values())
+
+    final_deps: list[str] = []
+    print("Checking dependencies against HPC module list...", file=sys.stderr)
+    for dep in deps:
+        name = normalise(pkg_name(dep))
+        # Smart check: if the package name is a substring of any loaded module name
+        # (e.g. 'torch' in 'pytorch', 'scipy' in 'scipy-bundle')
+        if any(name in mod for mod in module_names):
+            print(f"  [skip – module provider found] {dep}", file=sys.stderr)
+            continue
+
+        final_deps.append(dep)
+        print(f"  [pip] {dep}", file=sys.stderr)
+
+    hpc_dir = ROOT / "env" / "hpc"
+    output_path = hpc_dir / "requirements.txt"
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+    output_path.write_text("\n".join(final_deps) + "\n")
+    print(f"\nWrote {len(final_deps)} requirement(s) to {output_path}", file=sys.stderr)
+
+
+if __name__ == "__main__":
+    main()

scripts/hpc/install.sh

@@ -0,0 +1,54 @@
+#!/bin/bash -l
+# scripts/hpc/install.sh
+#
+# Usage (on any compute node):
+#   bash scripts/hpc/install.sh
+#
+# Batch usage:
+#   qsub scripts/hpc/install.sh
+
+#PBS -N brittlestar-install
+#PBS -l walltime=00:15:00
+
+set -euo pipefail
+
+# Preliminary status echo
+echo ">>> Starting installation job $PBS_JOBID on $(hostname)..."
+
+if [ -n "$PBS_O_WORKDIR" ]; then
+    cd "$PBS_O_WORKDIR"
+fi
+
+mkdir "${PBS_O_WORKDIR}/runs"
+
+# Mirror configs to $VSC_DATA to avoid home quota limits (3GB)
+# vsc-venv manages environments relative to the requirements file
+PROJ_NAME=$(basename "$PWD")
+HPC_CONFIG_DIR="$VSC_DATA/$PROJ_NAME/env/hpc"
+mkdir -p "$HPC_CONFIG_DIR"
+cp env/hpc/*.txt "$HPC_CONFIG_DIR/"
+
+# Keep caches off $VSC_HOME (quota ~3 GB).
+export PIP_CACHE_DIR="$VSC_SCRATCH/.cache/pip"
+export UV_CACHE_DIR="$VSC_SCRATCH/.cache/uv"
+mkdir -p "$PIP_CACHE_DIR" "$UV_CACHE_DIR"
+
+module load vsc-venv
+
+echo ">>> Synchronizing and activating environment (vsc-venv)..."
+# cd to $VSC_DATA so vsc-venv creates its venvs/ directory there, not in $HOME.
+mkdir -p "$VSC_DATA/$PROJ_NAME"
+cd "$VSC_DATA/$PROJ_NAME"
+set +euo pipefail
+source vsc-venv --activate \
+    --modules "$HPC_CONFIG_DIR/modules.txt" \
+    --requirements "$HPC_CONFIG_DIR/requirements.txt"
+set -euo pipefail
+cd "$PBS_O_WORKDIR"
+
+echo '>>> Installing ipykernel...'
+CLUSTER_ID="${VSC_INSTITUTE_CLUSTER:-generic}"
+python -m ipykernel install --user --name="sel3_${CLUSTER_ID}" \
+    --display-name "SEL3 (${CLUSTER_ID})"
+
+echo '>>> Done'

scripts/hpc/train.pbs

@@ -0,0 +1,63 @@
+# Production training (requires GPU at runtime):
+#   qsub -l gpus=1 scripts/hpc/train.pbs
+# Debug/CPU training:
+#   qsub scripts/hpc/train.pbs
+
+#PBS -N brittlestar-ppo
+#PBS -l nodes=1:ppn=8
+#PBS -l walltime=24:00:00
+#PBS -o runs/brittlestar-ppo.o$PBS_JOBID
+#PBS -e runs/brittlestar-ppo.e$PBS_JOBID
+
+set -euo pipefail
+
+# Preliminary status echo
+echo ">>> Starting training job $PBS_JOBID on $(hostname)..."
+
+if [ -n "$PBS_O_WORKDIR" ]; then
+    cd "$PBS_O_WORKDIR"
+fi
+
+# Set up storage paths dynamically
+PROJ_NAME=$(basename "$PWD")
+RUN_ID="brittlestar_${PBS_JOBID}"
+SCRATCH_RUNDIR="$VSC_SCRATCH/runs/$RUN_ID"
+DATA_RUNDIR="$VSC_DATA/runs/$RUN_ID"
+mkdir -p "$SCRATCH_RUNDIR" "$DATA_RUNDIR" runs/
+
+# Keep caches off $VSC_HOME (quota ~3 GB).
+export PIP_CACHE_DIR="$VSC_SCRATCH/.cache/pip"
+export UV_CACHE_DIR="$VSC_SCRATCH/.cache/uv"
+mkdir -p "$PIP_CACHE_DIR" "$UV_CACHE_DIR"
+
+module load vsc-venv
+
+echo ">>> Synchronizing and activating environment (vsc-venv)..."
+HPC_CONFIG_DIR="$VSC_DATA/$PROJ_NAME/env/hpc"
+if [ ! -d "$HPC_CONFIG_DIR" ]; then
+    echo "ERROR: HPC_CONFIG_DIR ($HPC_CONFIG_DIR) does not exist. Run install.sh first."
+    exit 1
+fi
+
+# cd to $VSC_DATA so vsc-venv finds its venvs/ directory there, not in $HOME.
+cd "$VSC_DATA/$PROJ_NAME"
+set +euo pipefail
+source vsc-venv --activate \
+    --modules "$HPC_CONFIG_DIR/modules.txt" \
+    --requirements "$HPC_CONFIG_DIR/requirements.txt"
+set -euo pipefail
+cd "$PBS_O_WORKDIR"
+
+
+echo ">>> Starting BrittleStar training..."
+export MUJOCO_GL=egl
+export WANDB_DIR="$SCRATCH_RUNDIR"
+
+python src/train.py \
+    --env-config-path configs/hpc/smoke_test.yaml \
+    --run-dir "$SCRATCH_RUNDIR"
+
+echo ">>> Staging out results to $DATA_RUNDIR..."
+cp -r "$SCRATCH_RUNDIR/." "$DATA_RUNDIR/"
+
+echo ">>> Done"

src/brittle_star_project/dataclasses/PPOArgs.py

@@ -11,11 +11,17 @@ class PPOArgs:
     """
 
     # path to environment config file, if None, use default config
-    config_path: str | None = None
+    env_config_path: str | None = None
 
     # the name of this experiment
     exp_name: str = "brittle_star_ppo"
 
+    # the directory to save the experiment results
+    run_dir: str | None = None
+
+    # how often to save checkpoints (0 to disable)
+    checkpoint_frequency: int = 0
+
     # seed of the experiment
     seed: int = 1
 

src/brittle_star_project/environment/BrittleStarJaxEnvWrapper.py

@@ -8,7 +8,7 @@ from brittle_star_project import (
     ArenaConfig,
     Backend,
 )
-from brittle_star_project.environment import from_json
+from brittle_star_project.environment import from_file
 
 
 class BrittleStarJaxEnvWrapper:
@@ -82,7 +82,7 @@ class BrittleStarJaxEnvWrapper:
     def from_config(
         config_path: str, num_envs: int, backend: Backend = Backend.MJX
     ) -> "BrittleStarJaxEnvWrapper":
-        morphology_cfg, arena_cfg, env_cfg = from_json(config_path)
+        morphology_cfg, arena_cfg, env_cfg = from_file(config_path)
         return BrittleStarJaxEnvWrapper(
             morphology_cfg, arena_cfg, env_cfg, num_envs=num_envs, backend=backend
         )

src/brittle_star_project/environment/__init__.py

@@ -1,4 +1,4 @@
-from .env_config import ArenaConfig, EnvConfig, MorphologyConfig, from_json
+from .env_config import ArenaConfig, EnvConfig, MorphologyConfig, from_file
 from .env_types import Backend, Task
 from .env_wrapper import BrittleStarEnv, StepResult
 from .factory import BrittleStarEnvFactory
@@ -12,5 +12,5 @@ __all__ = [
     "BrittleStarEnv",
     "StepResult",
     "BrittleStarEnvFactory",
-    "from_json",
+    "from_file",
 ]

src/brittle_star_project/environment/env_config.py

@@ -50,10 +50,17 @@ class EnvConfig:
     light_perlin_noise_scale: int = 0
 
 
-def from_json(path: str) -> tuple[MorphologyConfig, ArenaConfig, EnvConfig]:
+def from_file(path: str) -> tuple[MorphologyConfig, ArenaConfig, EnvConfig]:
+    """Load configurations from a JSON or YAML file."""
     with open(path, "r") as f:
-        config_json = json.load(f)
+        if path.endswith(".yaml") or path.endswith(".yml"):
-        morphology = MorphologyConfig(**config_json.get("morphology", {}))
+            import yaml
-        arena = ArenaConfig(**config_json.get("arena", {}))
+
-        env = EnvConfig(**config_json.get("env", {}))
+            config_dict = yaml.safe_load(f)
+        else:
+            config_dict = json.load(f)
+
+        morphology = MorphologyConfig(**config_dict.get("morphology", {}))
+        arena = ArenaConfig(**config_dict.get("arena", {}))
+        env = EnvConfig(**config_dict.get("env", {}))
         return morphology, arena, env

uv.lock

@@ -22,11 +22,16 @@ dependencies = [
     { name = "jax" },
     { name = "matplotlib" },
     { name = "mediapy" },
+    { name = "mujoco-warp" },
+    { name = "numpy" },
     { name = "optax" },
+    { name = "protobuf" },
     { name = "pyopengl" },
     { name = "pyopengl-accelerate" },
+    { name = "torch" },
     { name = "tyro" },
     { name = "wandb" },
+    { name = "warp-lang" },
 ]
 
 [package.optional-dependencies]
@@ -53,11 +58,16 @@ requires-dist = [
     { name = "jax", extras = ["cuda13"], marker = "extra == 'cuda'", specifier = "==0.9.0.1" },
     { name = "matplotlib", specifier = "==3.10.8" },
     { name = "mediapy", specifier = "==1.2.6" },
+    { name = "mujoco-warp" },
+    { name = "numpy", specifier = ">=2.0.0" },
     { name = "optax", specifier = ">=0.2.6" },
+    { name = "protobuf", specifier = ">=5.0.0" },
     { name = "pyopengl", specifier = ">=3.1.10" },
     { name = "pyopengl-accelerate", specifier = ">=3.1.10" },
+    { name = "torch", specifier = ">=2.4.0" },
     { name = "tyro", specifier = ">=1.0.10" },
     { name = "wandb", specifier = "==0.24.2" },
+    { name = "warp-lang" },
 ]
 provides-extras = ["cuda"]
 
@@ -1262,6 +1272,22 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/c2/7c/ad82beb7c4c9186d9fbef4799109d799692d70276bb1b3ee18a0674170d8/mujoco_mjx-3.6.0-py3-none-any.whl", hash = "sha256:c81000af0653f162b76009f48c153e9e6d19bfa8febe851e12466c81cbb7336a", size = 7013366, upload-time = "2026-03-11T01:46:19.148Z" },
 ]
 
+[[package]]
+name = "mujoco-warp"
+version = "3.6.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "absl-py" },
+    { name = "etils", extra = ["epath"] },
+    { name = "mujoco" },
+    { name = "numpy" },
+    { name = "warp-lang" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/0e/de/b853418268e9777cad2792ee3a145c8397e3d4517d136499645847ffd7f2/mujoco_warp-3.6.0.tar.gz", hash = "sha256:3c4111a4e13dc61268ddac52593ac5032c05a7d80f0c5e3c98bf5881e32b5d06", size = 1887269, upload-time = "2026-03-11T01:11:44.343Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/a3/b5/06c1e23c0cc4a06da268aa5f0fe05348d89c9ecbd3ecfb4c6d2b14ea23b2/mujoco_warp-3.6.0-py3-none-any.whl", hash = "sha256:371a405b186332cbfaa9630aabf35967405ef847cb7ea7c018ec67426a2ea160", size = 1965960, upload-time = "2026-03-11T01:11:42.527Z" },
+]
+
 [[package]]
 name = "mypy-extensions"
 version = "1.1.0"
@@ -1630,7 +1656,7 @@ name = "pexpect"
 version = "4.9.0"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
-    { name = "ptyprocess" },
+    { name = "ptyprocess", marker = "sys_platform != 'emscripten' and sys_platform != 'win32'" },
 ]
 sdist = { url = "https://files.pythonhosted.org/packages/42/92/cc564bf6381ff43ce1f4d06852fc19a2f11d180f23dc32d9588bee2f149d/pexpect-4.9.0.tar.gz", hash = "sha256:ee7d41123f3c9911050ea2c2dac107568dc43b2d3b0c7557a33212c398ead30f", size = 166450, upload-time = "2023-11-25T09:07:26.339Z" }
 wheels = [
@@ -2428,6 +2454,20 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/3a/9a/f3919d7ee7ba99dabf0aac7e299c6c328f5eae94f9f6b28c76005f882d5d/wandb-0.24.2-py3-none-win_arm64.whl", hash = "sha256:b42614b99f8b9af69f88c15a84283a973c8cd5750e9c4752aa3ce21f13dbac9a", size = 20268261, upload-time = "2026-02-05T00:12:14.353Z" },
 ]
 
+[[package]]
+name = "warp-lang"
+version = "1.12.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "numpy" },
+]
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/7d/15/fadf3e3ba5c1c907530c20c98402aaef792da74bbbe382c848cef6e5affe/warp_lang-1.12.0-py3-none-macosx_11_0_arm64.whl", hash = "sha256:c78c3701d5cad86c30ef5017410d294ec46a396bb0d502ee1c98743494f3a62f", size = 24168341, upload-time = "2026-03-06T19:42:16.333Z" },
+    { url = "https://files.pythonhosted.org/packages/98/13/deab9dbae5c6aa753ac8ea1d3b1f85d20c5bab7bdebd8916ce242fbe1f0b/warp_lang-1.12.0-py3-none-manylinux_2_28_x86_64.whl", hash = "sha256:a1436f60a1881cd94f787e751a83fc0987626be2d3e2b4e74c64a6947c6d1266", size = 136485344, upload-time = "2026-03-06T19:43:02.427Z" },
+    { url = "https://files.pythonhosted.org/packages/45/ce/9f5c57cac849edaba2f3335cb649b7019b09195b3af02221258482254559/warp_lang-1.12.0-py3-none-manylinux_2_34_aarch64.whl", hash = "sha256:a2d6decba693aba5b828573c4414fd6a3f4c4a934db9c322736ef2b3fa99fe76", size = 137735580, upload-time = "2026-03-06T19:44:22.279Z" },
+    { url = "https://files.pythonhosted.org/packages/7e/3f/1ddc888fe769447ae33915a9567a9dd7467e1fc7fc8010d39e01b339667f/warp_lang-1.12.0-py3-none-win_amd64.whl", hash = "sha256:697248edd2f1e2952f50e3db33b214af76173641a8894aacc467bed6dc247f8a", size = 119793582, upload-time = "2026-03-06T19:45:37.288Z" },
+]
+
 [[package]]
 name = "wcwidth"
 version = "0.6.0"