BEAVR Sim - GPU-Accelerated RL & Imitation Learning

Key Features

Capable of running parallel environments on a single GPU
Unified API for observations and actions via LeRobot
YAML-based robot configurations, which allow easy use of the entire MuJoCo Menagerie library
Flexible task and scene definitions allowing scene building (as seen below)
Support for RL and Imitation Learning on any LeRobot-compatible environment
4 pre-built scenes with 100 human demonstrations each collected via teleoperation using BEAVR
Integration with Isaac Lab APIs
Fully compatible with LeRobot APIs for training and evaluation

Demonstrations

Below are clips from datasets we collected with the BEAVR teleoperation system. Each dataset contains 100 demonstrations from four memory-based manipulation tasks.

Pick and Place

Simple manipulation task demonstrating object grasping and placement. The system learns to pick up objects from randomized positions and place them at a target location,

Manipulation Grasping IL

Shell Game

Advanced memory/tracking and manipulation task where the robot must track the cup containing a concealed object. This scene is designed to test object permanence and memory amid occlusions through imitation learning.

Tracking Memory IL

Server Swap - Overhead View

Top-down perspective of a mobile manipulation task simulating server module replacement. The robot coordinates its arm and mobile base to insert server components in the correct slot indicated by a transient LED (solid orange for 5s). This scene is complex, requiring manipulation, memory of the target slot, and movement across the scene.

Egocentric Precision Assembly

Server Swap - Egocentric View

First-person perspective of the server swap task, showing the detailed interaction from the robot's viewpoint.

Egocentric Precision Assembly

Vanishing Blueprint

Manipulation and memory task testing the robot's ability to recall the correct order of objects for stacking. A blueprint is shown in the first 5s before vanishing, requiring the robot to remember the sequence while stacking.

Stacking Memory Configuration

Citations

BEAVR-teleop

@misc{posadasnava2025beavr,
  title         = {BEAVR: Bimanual, multi-Embodiment, Accessible, Virtual Reality Teleoperation System for Robots},
  author        = {Alejandro Posadas-Nava and Alejandro Carrasco and Richard Linares},
  year          = {2025},
  eprint        = {2508.09606},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  note          = {Accepted for presentation at ICCR 2025, Kyoto},
  url           = {https://arxiv.org/abs/2508.09606}
}

https://github.com/ARCLab-MIT/beavr-bot

MuJoCo Menagerie

@software{menagerie2022github,
  author = {Zakka, Kevin and Tassa, Yuval and {MuJoCo Menagerie Contributors}},
  title = {{MuJoCo Menagerie: A collection of high-quality simulation models for MuJoCo}},
  url = {https://github.com/google-deepmind/mujoco_menagerie},
  year = {2022}
}

https://github.com/google-deepmind/mujoco_menagerie

LeRobot

@misc{cadene2024lerobot,
  author = {Cadene, Remi and Alibert, Simon and Soare, Alexander and Gallouedec, Quentin and Zouitine, Adil and Palma, Steven and Kooijmans, Pepijn and Aractingi, Michel and Shukor, Mustafa and Aubakir, Riyad and Charraut, Thomas and Liu, Shengxiang and Bogdan, Cosmin and Herdt, Kaat and Schoukens, Sanne and Sferrazza, Carmelo and Wassing, Nick and Taiana, Mariano and Coumans, Erwin and Clegg, Alexander and Strudel, Robin and Jeanneret, Robin},
  title = {LeRobot: State-of-the-art Machine Learning for Real-World Robotics in Pytorch},
  howpublished = "\url{https://github.com/huggingface/lerobot}",
  year = {2024}
}

https://github.com/huggingface/lerobot

MJLab

@software{Zakka_mjlab_Isaac_Lab_2025,
  author = {Zakka, Kevin and Yi, Brent and Liao, Qiayuan and Le Lay, Louis},
  license = {Apache-2.0},
  month = dec,
  title = {{mjlab: Isaac Lab API, powered by MuJoCo-Warp, for RL and robotics research.}},
  url = {https://github.com/mujocolab/mjlab},
  version = {1.0.0},
  year = {2025}
}

https://github.com/mujocolab/mjlab