praxis-eval

praxis-eval is a standalone robot-policy evaluation package for simulation benchmarks. It owns benchmark setup, rollout execution, metrics, artifacts, and observation/action contracts; users provide a policy adapter that consumes documented observations and returns documented actions.

Installation

praxis-eval supports Python 3.10 and newer.

pip install praxis-eval

Install the benchmark extras you plan to run:

pip install "praxis-eval[libero]"
pip install "praxis-eval[robocasa]"
pip install "praxis-eval[robocasa_gr1]"
pip install "praxis-eval[robomimic]"
pip install "praxis-eval[metaworld]"
pip install "praxis-eval[simpler]"
pip install "praxis-eval[mshab]"
pip install "praxis-eval[remote]"

Minimal Local Evaluation

This example assumes the LIBERO extra and simulator runtime are installed.

import numpy as np

from praxis_eval import EvalConfig, evaluate, get_driver


class RandomPolicy:
    def __init__(self, seed: int = 0) -> None:
        self.rng = np.random.default_rng(seed)

    def reset(self, episode_ids=None) -> None:
        pass

    def act(self, observations, *, action_spec=None, policy_kwargs=None, episode_ids=None):
        del policy_kwargs, episode_ids
        if action_spec is None or action_spec.shape is None:
            raise ValueError("This example needs a benchmark ActionSpec.")
        low = -1.0 if action_spec.minimum is None else action_spec.minimum
        high = 1.0 if action_spec.maximum is None else action_spec.maximum
        return self.rng.uniform(
            low=low,
            high=high,
            size=(len(observations), *action_spec.shape),
        ).astype(action_spec.dtype)


driver = get_driver("libero")
print(driver.contract)

result = evaluate(
    "libero",
    policy=RandomPolicy(seed=42),
    config=EvalConfig(
        task="libero_10",
        task_ids=(0,),
        num_eval_per_task=1,
        num_parallel_env=1,
        output_dir="eval/libero_smoke",
    ),
)

print(result.overall)
print(result.artifacts)

Supported Benchmarks

Use the setup and verification CLIs to inspect available benchmark-specific commands:

praxis-eval-setup --help
praxis-eval-verify --help

Documentation

Full documentation is available at https://chaoqi-liu.com/praxis-eval/, including installation details, benchmark contracts, remote policy evaluation, runtime setup, examples, and the developer guide for adding new benchmarks.

Examples

Runnable examples live in examples/. They cover local policy evaluation, remote policy evaluation, custom policy adapters, and custom benchmark drivers.

License And Citation

praxis-eval is licensed under Apache-2.0. If this package supports your research or product work, cite the repository using CITATION.cff.