XPolicyLab/XPolicyLab

What it solves

XPolicyLab provides a unified standard and open ecosystem for evaluating and deploying robot policies. It solves the problem of fragmented evaluation pipelines by acting as a shared layer between policy code and evaluation environments, reducing the complexity of connecting N policies to M environments from $O(N \times M)$ to $O(N + M)$.

How it works

The framework separates model-side dependencies from environment-side dependencies, allowing each to run in its own isolated environment (e.g., separate conda or uv environments) and communicate via a websocket-based policy server.

Key components include:

• Policy Adapters: Each policy has a dedicated adapter containing a model.py (implementing a standardized API for observations and actions) and a deploy.py bridge.
• Policy Server: Handles the serving of the model, observation decoding, and communication with the environment client.
• Environment Client: A benchmark, simulator, or real-robot setup that plugs into the policy-side interface.

Who it’s for

It is designed for robotics researchers and developers who need to evaluate robot policies across multiple benchmarks (such as RoboDojo and RoboTwin) or deploy them to physical robots without coupling the policy code to a specific environment.

Highlights

• Large Policy Zoo: Integrates 49 robot policies, including VLA, world-action, imitation-learning, and memory-augmented families.
• Environment Isolation: Allows policies to run in their own dependencies while the simulator or robot client runs separately.
• Remote Deployment: Supports connecting policy servers and environment clients across different machines via websockets.
• Benchmark Agnostic: Integrates with major benchmarks like RoboDojo, RoboTwin, and RMBench.
• Standardized Contract: Provides a common lifecycle for installation, data conversion, training, and evaluation.