takuseno/d3rlpy

Health v2 · maintenance only

Adoption not part of the score

1684 stars · 268 forks observed · 2026-09-30

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

d3rlpy is a Python library for offline and online deep reinforcement learning built on PyTorch, offering state-of-the-art algorithms through a scikit-learn style API. It supports distributional Q functions, data-parallel distributed training, and works with datasets from D4RL, PyBullet, and Atari environments.

Use cases

• train RL policies from logged datasets without online interaction
• run offline RL experiments on D4RL benchmark datasets
• train online RL agents on gymnasium environments
• scale offline RL training across multiple GPUs or nodes
• reproduce results from offline RL research papers
• apply RL to robotics or medical tasks where online exploration is infeasible

When to choose

• you need offline RL algorithms like CQL with a simple, well-documented API
• you want both offline and online RL in one library
• you prefer scikit-learn style APIs over writing raw PyTorch training loops
• you need distributed or GPU-accelerated RL training

When to avoid

• you need a lightweight tabular RL library rather than deep RL
• you require environments or algorithms outside its supported set
• you are locked to an older version and cannot handle v2.x breaking changes

Facets

library · maturity active

machine-learning deep-learning reinforcement-learning reinforcement-learning machine-learning deep-learning python windows offline-rl pytorch scikit-learn-style-api distributional-q-function distributed-training linux macos gpu

Member repositories

For agents

markdown · JSON · MCP: product_card(name="takuseno/d3rlpy")