ADARL - Adapters for Robot Learning

ADARL's core feature is a collection of interfaces (i.e. Adapters) that sit between Reinforcement Learning environments and simulations or real world robotics frameworks.

ADARL allows you to write an environment once against a backend-agnostic adapter interface, and then run it unchanged on different simulators (or on real hardware) — switching backend is just a matter of swapping the adapter. This makes it easy to, for example, train in a fast GPU-batched simulator and deploy the same environment on a CPU simulator or a physical robot.

Adapters

Adapters live in adarl/adapters and are organized as a small capability-based interface hierarchy rather than a single monolithic API:

• Core: BaseVecAdapter provides the common lifecycle (build scenario, reset, step, read joint/link state, render) over a batch of parallel simulations.
• Control modes are separate mixin interfaces, so a backend only implements what it supports: BaseVecJointImpedanceAdapter, BaseVecJointEffortAdapter, BaseVecJointPositionAdapter.
• Simulation-only capabilities (setting state directly, spawning models, configuring collision pairs) live in BaseVecSimulationAdapter.

Concrete backends:

Adapter	Backend	Notes
MjxAdapter, MjxJointImpedanceAdapter, MjxActuatedAdapter	MuJoCo MJX	GPU-batched, JAX/XLA/Warp
MujocoAdapter, MujocoJointImpedanceAdapter	MuJoCo classic	CPU, single simulation (vec_size == 1)
PyBulletAdapter, PyBulletJointImpedanceAdapter, PyBullet2DofCartesianAdapter	PyBullet	CPU
ZmqXbotAdapter / VecZmqXbotAdapter, StandaloneRealAdapter	Real robots	hardware deployment (e.g. XBot over ZMQ)

VecSimJointImpedanceAdapterWrapper adapts a single-simulation adapter to the vectorized (Vec) interface, so single-sim backends can be used where a batched adapter is expected.

Environment structure

Environments are vectorized: they operate on a batch of robot systems at once. The environment definition framework lives in adarl/envs/vec:

• Env logic — BaseVecEnv → ControlledVecEnv define how observations, rewards and resets are computed for the batch, driving the simulation through an adapter.
• Runner — EnvRunner drives the env: stepping, reset / autoreset, episode bookkeeping and logging. Wrappers like EnvRunnerRecorderWrapper add video/info recording on top of a runner.
• Gym exposure — Runner2GymWrapper and Runner2VecGymWrapper expose a runner as a standard Gymnasium env / vector env, so trained agents and existing RL tooling can consume it directly.

(A set of older single-environment classes also exists under adarl/envs, but are kept only for legacy reasons.)

Repository layout

• adarl/adapters — the adapter interfaces and backend implementations.
• adarl/envs/vec — the vectorized environment framework (env base classes, runner, Gym wrappers).
• adarl/utils — supporting utilities (logging, tensor-tree helpers, run/session management, debugging).
• adarl/models, adarl/assets — robot/scene models and assets.
• adarl/examples — runnable examples (e.g. cartpole, ant, half-cheetah).