SLM Lab

Modular Deep Reinforcement Learning framework in PyTorch.
Companion library of the book Foundations of Deep Reinforcement Learning.
Documentation · Benchmark Results

NOTE: v5.0 updates to Gymnasium, uv tooling, and modern dependencies with ARM support - see CHANGELOG.md.

Book readers: git checkout v4.1.1 for Foundations of Deep Reinforcement Learning code.

BeamRider	Breakout	KungFuMaster	MsPacman
Pong	Qbert	Seaquest	Sp.Invaders
Ant	HalfCheetah	Hopper	Humanoid
Inv.DoublePendulum	InvertedPendulum	Reacher	Walker

SLM Lab is a software framework for reinforcement learning (RL) research and application in PyTorch. RL trains agents to make decisions by learning from trial and error—like teaching a robot to walk or an AI to play games.

What SLM Lab Offers

Feature	Description
Ready-to-use algorithms	PPO, SAC, DQN, A2C, REINFORCE—validated on 70+ environments
Easy configuration	JSON spec files fully define experiments—no code changes needed
Reproducibility	Every run saves its spec + git SHA for exact reproduction
Automatic analysis	Training curves, metrics, and TensorBoard logging out of the box
Cloud integration	dstack for GPU training, HuggingFace for sharing results

Algorithms

Algorithm	Type	Best For	Validated Environments
REINFORCE	On-policy	Learning/teaching	Classic
SARSA	On-policy	Tabular-like	Classic
DQN/DDQN+PER	Off-policy	Discrete actions	Classic, Box2D, Atari
A2C	On-policy	Fast iteration	Classic, Box2D, Atari
PPO	On-policy	General purpose	Classic, Box2D, MuJoCo (11), Atari (54)
SAC	Off-policy	Continuous control	Classic, Box2D, MuJoCo

See Benchmark Results for detailed performance data.

Environments

SLM Lab uses Gymnasium (the maintained fork of OpenAI Gym):

Category	Examples	Difficulty	Docs
Classic Control	CartPole, Pendulum, Acrobot	Easy	Gymnasium Classic
Box2D	LunarLander, BipedalWalker	Medium	Gymnasium Box2D
MuJoCo	Hopper, HalfCheetah, Humanoid	Hard	Gymnasium MuJoCo
Atari	Breakout, MsPacman, and 54 more	Varied	ALE

Any gymnasium-compatible environment works—just specify its name in the spec.

Quick Start

# Install
uv sync
uv tool install --editable .

# Run demo (PPO CartPole)
slm-lab run                                    # PPO CartPole
slm-lab run --render                           # with visualization

# Run custom experiment
slm-lab run spec.json spec_name train          # local training
slm-lab run-remote spec.json spec_name train   # cloud training (dstack)

# Help (CLI uses Typer)
slm-lab --help                                 # list all commands
slm-lab run --help                             # options for run command

# Troubleshoot: if slm-lab not found, use uv run
uv run slm-lab run

Cloud Training (dstack)

Run experiments on cloud GPUs with automatic result sync to HuggingFace.

# Setup
cp .env.example .env  # Add HF_TOKEN
uv tool install dstack  # Install dstack CLI
# Configure dstack server - see https://dstack.ai/docs/quickstart

# Run on cloud
slm-lab run-remote spec.json spec_name train           # CPU training (default)
slm-lab run-remote spec.json spec_name search          # CPU ASHA search (default)
slm-lab run-remote --gpu spec.json spec_name train     # GPU training (for image envs)

# Sync results
slm-lab pull spec_name    # Download from HuggingFace
slm-lab list              # List available experiments

Config options in .dstack/: run-gpu-train.yml, run-gpu-search.yml, run-cpu-train.yml, run-cpu-search.yml

Minimal Install (Orchestration Only)

For a lightweight box that only dispatches dstack runs, syncs results, and generates plots (no local ML training):

uv sync --no-default-groups
uv run --no-default-groups slm-lab run-remote spec.json spec_name train
uv run --no-default-groups slm-lab pull spec_name
uv run --no-default-groups slm-lab plot -f folder1,folder2

Citation

If you use SLM Lab in your research, please cite:

@misc{kenggraesser2017slmlab,
    author = {Keng, Wah Loon and Graesser, Laura},
    title = {SLM Lab},
    year = {2017},
    publisher = {GitHub},
    journal = {GitHub repository},
    howpublished = {\url{https://github.com/kengz/SLM-Lab}},
}

License

MIT