Changelog

This page records public releases of the rlmesh Python package. For API stability and peer version rules, see Compatibility.

0.1.0

RLMesh 0.1.0 is the first public release. It connects models to Gymnasium-style environments for local and remote evaluation. The quickstart walks through a first run.

Environments and models

• Author environments with EnvFactory and serve them with python -m rlmesh.serve. Use EnvServer to serve an existing environment object. RemoteEnv and RemoteVectorEnv connect clients to those endpoints.
• Run models locally or against remote and sandboxed environments with Model, RemoteModel, SandboxModel, and run(). Use Session when you need direct control over reset, predict, and step calls.
• Serve several independent environment lanes from one endpoint. The runtime handles each lane’s episode lifecycle and groups predictions when a model implements a compatible batched method.
• Blocking environment serving runs simulator calls on the thread that built the environment. A vector environment with one lane serves as a scalar endpoint, including per-episode seeds and step caps. See the Isaac Sim example.
• Implement single action, action chunk, batched, or batched chunk prediction. The runtime handles chunk replay and execution horizons. Optional prefetch_lead overlaps prediction with chunk execution; prefetched actions can use an earlier observation than synchronous evaluation.
• Configure a model subclass with from_config(**config). A caller-owned model or environment remains open after a run; RLMesh closes instances it creates from a class or factory.

See Environment Reference, Model Reference, and Run an Evaluation for the full interfaces.

Model and environment data

• Declare environment outputs with tags and model inputs and actions with specs. The adapter resolves matching roles and reports its conversions with adapter.explain() and adapter.advisories().
• Name repeated body parts, joint axes, coordinate frames, and state provenance in the spec. Adapters can align labeled axes, convert supported image and rotation encodings, stack image frames, and read a model’s previous action.
• Use NumPy, Torch, or JAX value backends. rlmesh.Tensor transports shape, dtype, and bytes across the boundary.

See Adapters, Adapter Reference, and Value Backends.

Evaluation tools

• Set an exact episodes budget and supply seeds or trial indices for repeatable runs. RunHooks observes step and episode events on native and session runs. EpisodeResult carries per-episode timings, while a native RunResult includes aggregate telemetry. success_rate uses only outcomes the environment reports and is None when any outcome is unknown.
• Use the terminal or HTTP viewer during a run, or export results with rlmesh.Recorder as an rlmesh.result.v1 bundle.
• Build or run environments and models in isolated containers with the sandbox APIs. describe() and the image check command inspect package metadata before submission.

See Run an Evaluation, Sandbox Environments, and Performance and Scaling.

Managed platform and compatibility

• Sign in, select an organization, authenticate to the registry, and submit evaluations through the CLI. The experimental rlmesh.platform.Client calls the managed platform from Python. Managed access is in early access.
• The release uses the 2026.06 workflow edition and rlmesh-wire-v1 protocol generation. An environment or model can declare its edition in source; the runtime selects and pins a session edition supported by every participant. Managed images must use the SDK version expected by the platform runner.

See Managed Platform, Workflow Editions, and Compatibility.