Choose a training framework

Choose a training framework#

EmbodiChain supports RLinf and LeRobot integrations as well as native training. Choose the workflow by its environment/data contract and checkpoint format.

Workflow

Training owner

Start here

Native RL with PPO, GRPO or differentiable APG

EmbodiChain RL

Reinforcement Learning

External RL workers consuming EmbodiChain Gym tasks

RLinf

Using EmbodiChain with RLinf

Visual imitation learning with official ACT

LeRobot 0.6.1

Using EmbodiChain with LeRobot

Model-owned supervised objectives such as dynamics learning

EmbodiChain SL

Supervised Learning

Shared simulation and data, separate training runtimes#

EmbodiChain owns task registration, scene/robot configuration, sensors, observations, action semantics, rewards and physical evaluation. RLinf owns its distributed actors, rollouts and trainer. LeRobot owns ACT training, processors, normalization and native checkpoints. Installing EmbodiChain does not install RLinf; follow its linked integration recipe in the worker environment.

embodichain train-rl and embodichain train-sl launch native trainers. The external workflows use their own launchers; their checkpoints are not interchangeable with native trainer checkpoints. Use the ACT example’s evaluator for its portable LeRobot policies and the RLinf workflow for RLinf artifacts.

Before running an experiment#

  1. Choose a runnable task deployment and preserve its file-owned physics backend.

  2. Match observation features, units, control-part ordering, action semantics and control cadence to the policy. A matching tensor shape alone is insufficient.

  3. Install the selected dependency profile in every process that consumes the environment or data. See Installation and the framework guide.

  4. Run a small collection/training/evaluation smoke workflow before scaling.

  5. Preserve task configuration, dependency versions, dataset provenance and checkpoint processors. Keep training, development and test scenes separate.

LeRobot dataset compatibility does not imply support for every LeRobot policy. Likewise, RLinf task discovery does not prove that a particular VLA observation or chunked-action mapping is implemented. The linked guides describe the concrete workflows and adaptation boundaries.

Contents