Skip to content
RobothecaBeta
Log in

TorchRL

  • Software
  • Humanoids and robot learning
  • Library

TorchRL is a PyTorch-native toolkit of composable components for reinforcement learning, decision making, robotics and simulation, built around named tensor-dictionary data flow.

Made by PyTorch.

Source on GitHub: github.com/pytorch/rl (opens in a new tab). Facts read at commit e7a05a7 (opens in a new tab).

Licenses

Code
MIT (opens in a new tab)

Read by Robotheca from LICENSE (opens in a new tab) (MIT) at commit e7a05a7 (opens in a new tab), on 16 Oct 2026.

Commercial use
Yes

What it does

  • Provides a TensorDict-based data model so observations, actions, rewards and recurrent states keep names, structure and devices through the training loop.
  • Offers environments, policy wrappers, transforms, collectors, replay buffers and loss modules that can be swapped independently.
  • Includes native and wrapped environments (Gymnasium, DM Control, PettingZoo, VMAS and others) with vectorized and multiprocess execution.
  • Ships collectors for single-process, async, multiprocess and distributed rollouts, plus modular and prioritized replay buffers.
  • Implements many objectives, from PPO and SAC to offline, imitation, model-based and multi-agent methods, with value-estimator utilities.
  • Provides example implementations and tutorials, and is installable from PyPI, with optional extras for specific environment libraries.

Written by Robotheca from the project's README (opens in a new tab) at commit e7a05a7.

Getting started

Difficulty
Intermediate

Maintenance

Status
Active: a commit in the last 90 days
Newest commit
16 Oct 2026 on main
Latest release
TorchRL v0.14.1 (10 Oct 2026)

Checked on GitHub on 16 Oct 2026.

Links

Robots that can run it

  • No photo yet
    Quadruped

    Physical Ant

    A quadruped ant-style robot with eight Dynamixel actuators that runs learned locomotion policies directly on hardware, controlled by a single computer over USB.

    • Works with AI
    Level
    Intermediate
    Buildability
    75/100
    Buildable with some effort
    Cost
    $2,230.61≈ €1,982 (ECB rate, 16 Oct 2026)

    Found by Robotheca in requirements.txt (opens in a new tab)

Robotheca lists software as a reference: nothing is hosted or run here, and every fact links to where it was read.