---

PolaRiS

🌐 Website • 📄 Paper • 🤗 PolaRiS Hub

PolaRiS is a evaluation framework for generalist policies. It provides tooling for reconstructing environments, evaluating models, and running experiments with minimal setup.

Installation

Clone the repository (recursively)

git clone --recursive [email protected]:arhanjain/polaris.git
cd PolaRiS

If you cloned without --recursive:

git submodule update --init --recursive

Setup environment with uv

If you don't have UV installed, see installation instructions

By default we support CUDA 13. If you have an older version of CUDA installed, please downgrade the torch and torchvision version and index to be compatible in the pyproject.toml.

uv sync 

Getting Started

First, download the PolaRiS environments (<2GB)

uvx hf download owhan/PolaRiS-Hub --repo-type=dataset --local-dir ./PolaRiS-Hub

Minimal Code Example

Next let's test a simple random action policy in a PolaRiS environment.

import torch
import argparse
import gymnasium as gym
from isaaclab.app import AppLauncher
# This must be done before importing anything with dependency on Isaaclab
# >>>> Isaac Sim App Launcher <<<<
parser = argparse.ArgumentParser()
args_cli, _ = parser.parse_known_args()
args_cli.enable_cameras = True
args_cli.headless = True
app_launcher = AppLauncher(args_cli)
simulation_app = app_launcher.app
# >>>> Isaac Sim App Launcher <<<<

import polaris.environments
from isaaclab_tasks.utils import parse_env_cfg  # noqa: E402
from polaris.environments.manager_based_rl_splat_environment import ManagerBasedRLSplatEnv
from polaris.utils import load_eval_initial_conditions

env_cfg = parse_env_cfg(
    "DROID-FoodBussing",
    device="cuda",
    num_envs=1,
    use_fabric=True,
)
env: ManagerBasedRLSplatEnv = gym.make("DROID-FoodBussing", cfg=env_cfg)   # type: ignore
language_instruction, initial_conditions = load_eval_initial_conditions(env.usd_file)
obs, info = env.reset(object_positions = initial_conditions[0])

while True:
    action = torch.tensor(env.action_space.sample())
    obs, rew, term, trunc, info = env.step(action, expensive=True)

    if term[0] or trunc[0]:
        break

print(f"Episode Finished. Success: {info['rubric']['success']}, Progress: {info['rubric']['progress']}")

Run a π0.5 Policy in PolaRiS

Note: First run may take longer due to JIT compilation of the splat rasterization kernels. Ensure you have NVIDIA Drivers and CUDA Toolkit (nvcc) properly configured.

Both the policy server and evaluation process should fit onto a single GPU (tested on RTX 3090, 24 GB).

# Starting from the root of this repo. This will setup openpi and host a pi05 policy.
cd third_party/openpi
GIT_LFS_SKIP_SMUDGE=1 uv sync
GIT_LFS_SKIP_SMUDGE=1 uv pip install -e .
XLA_PYTHON_CLIENT_MEM_FRACTION=0.35 uv run scripts/serve_policy.py --port 8000 policy:checkpoint --policy.config pi05_droid_jointpos_polaris --policy.dir gs://openpi-assets/checkpoints/polaris/pi05_droid_jointpos_polaris

# In a separate process, start evaluation process
sudo apt install ffmpeg # for saving videos
uv run scripts/eval.py --environment DROID-FoodBussing --policy.port 8000 --run-folder runs/test

Results include rollout videos, and a CSV summarizing success and normalized progress of each episode.

Off-the-shelf Evaluation Environments

Environment Name	Prompt	Image
DROID-BlockStackKitchen	Place and stack the blocks on top of the green tray
DROID-FoodBussing	Put all the foods in the bowl
DROID-PanClean	Use the yellow sponge to scrub the blue handle frying pan
DROID-MoveLatteCup	put the latte art cup on top of the cutting board
DROID-OrganizeTools	put the scissor into the large container
DROID-TapeIntoContainer	put the tape into the container

PolaRiS-Ready Policies

All checkpoints for PolaRiS were based on DROID base policies. Checkpoints were produced by cotraining at a weightage of 10% random simulated data and 90% DROID data for 1k steps.

Policy Name	Checkpoints Path
π0.5 Polaris	gs://openpi-assets/checkpoints/polaris/pi05_droid_jointpos_polaris
π0 Fast Polaris	gs://openpi-assets/checkpoints/polaris/pi0_fast_droid_jointpos_polaris
π0 Polaris	gs://openpi-assets/checkpoints/polaris/pi0_droid_jointpos_polaris
π0 Polaris (100k)	gs://openpi-assets/checkpoints/polaris/pi0_droid_jointpos_100k_polaris
PaliGemma Polaris	gs://openpi-assets/checkpoints/polaris/paligemma_binning_droid_jointpos_polaris

For the full list of all checkpoints, base policies, and environments we provide for evaluation, see checkpoints_and_envs.md

Cotraining and Evaluating Your Policies In PolaRiS

1. Download DROID simulated cotraining dataset
2. Cotrain a policy
   1. Using OpenPI
      • We provide cotraining configs for 4 policies in openpi
      • We already provide a client to inference openpi DROID policies in src/polaris/policy/droid_jointpos_client.py
   2. Training a custom policy
   • We recommend co-finetuning your policy with the provided sim dataset at 10% weightage
   • May need to define a custom policy client if your policy is not compatible with the provided DROID JointPosition client

See custom_policies.md for more details

Creating Custom Evaluation Environments

Time Estimate: 20 Minutes Human Time + 40 Minutes Offline Training

1. Take a video
2. Extract splat and mesh (we use 2DGS, but any method that produces both can work)
3. Compose environment USD using our provided Web GUI
4. Run evaluation :)
5. Contribute to the community pool of evaluation environments!

For detailed instructions, see docs/custom_environments.md

Issues

This codebase has been tested on CUDA 13 and CUDA 12 with NVIDIA 5090 and 3090 GPUs. Please raise an issue if you run into any issues.

Citation

If you find this repository useful, please consider citing it as:

@misc{jain2025polarisscalablerealtosimevaluations,
      title={PolaRiS: Scalable Real-to-Sim Evaluations for Generalist Robot Policies}, 
      author={Arhan Jain and Mingtong Zhang and Kanav Arora and William Chen and Marcel Torne and Muhammad Zubair Irshad and Sergey Zakharov and Yue Wang and Sergey Levine and Chelsea Finn and Wei-Chiu Ma and Dhruv Shah and Abhishek Gupta and Karl Pertsch},
      year={2025},
      eprint={2512.16881},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2512.16881}, 
}