PolaRiS Evaluation#

PolaRiS is a tabletop manipulation simulation platform with DROID-style tasks such as TapeIntoContainer and MoveLatteCup. RLinf supports evaluating OpenPI policies on PolaRiS.

Related training doc: RL with PolaRiS Simulation Platform

Environment Setup#

bash requirements/install.sh embodied --model openpi --env polaris
source .venv/bin/activate
export POLARIS_DATA_PATH=/path/to/dataset/PolaRiS-Hub

Example Configs#

Config file	Task	Model
`polaris_tapeintocontainer_openpi_pi05_eval.yaml`	TapeIntoContainer	π₀.₅
`polaris_movelattecup_openpi_eval.yaml`	MoveLatteCup	π₀

End-to-End Workflow#

Step 1: Download dataset and model

Follow RL with PolaRiS Simulation Platform to download the PolaRiS dataset and OpenPI checkpoints.

Step 2: Set environment variables

source .venv/bin/activate
export POLARIS_DATA_PATH=/path/to/dataset/PolaRiS-Hub

Step 3: Edit the config

Set rollout.model.model_path to your local checkpoint.

Step 4: Launch evaluation

bash evaluations/run_eval.sh polaris polaris_tapeintocontainer_openpi_pi05_eval

Or:

bash evaluations/run_eval.sh polaris polaris_movelattecup_openpi_eval

Step 5: Check results

The terminal prints eval/success_once; see Logs and Results for logs.

FAQ#

Dataset path: POLARIS_DATA_PATH must point to the PolaRiS-Hub root; run_eval.sh reads it automatically.
Model conversion: JAX checkpoints must be converted to PyTorch format per the training doc before evaluation.