MagicBot-VGA / CubeV2

MagicBot-VGA is a robot learning codebase for VLA policy training, simulation evaluation, and real-robot deployment. This repository contains the CubeV2 policy family, MagicBot_R0 experiments, RoboTwin/Libero evaluation helpers, and real-robot serving clients for Lift2 and Piper.

What Is Here

• src/lerobot/policies/cubev2: CubeV2 policy, transforms, and model code.
• src/lerobot/policies/MagicBot_R0: MagicBot_R0 policy and dataset pipeline.
• launch/: training and finetuning entrypoints.
• evaluation/RoboTwin: RoboTwin 2.0 evaluation workflow.
• evaluation/Real_Piper: sync real-robot Piper serving and ROS1 client.
• evaluation/Real_Lift2: Lift2 real-robot serving and inference runtime.
• evaluation/Libero: Libero evaluation helpers.
• util_scripts/: dataset conversion, norm-stat computation, checkpoint repacking, and submission packaging utilities.

Quick Setup

The core environment is tested with Python 3.10, CUDA 12.8, and PyTorch 2.7.1.

conda create -y -n magicbot python=3.10
conda activate magicbot
pip install --upgrade pip

conda install -c conda-forge ffmpeg=7.1.1 svt-av1 -y

pip install torch==2.7.1 torchvision==0.22.1 torchaudio==2.7.1 \
  --index-url https://download.pytorch.org/whl/cu128

pip install torchcodec numpy scipy transformers==4.57.1 mediapy loguru pytest omegaconf h5py
pip install -e .

For real-robot serving, also install:

pip install tyro matplotlib mediapy websockets msgpack

Qwen3-VL Patch

CubeV2 uses a patched Qwen3-VL implementation for cached inference. After installing transformers==4.57.1, copy the repository patch into the installed package:

TRANSFORMERS_DIR=${CONDA_PREFIX}/lib/python3.10/site-packages/transformers/
cp -r src/lerobot/policies/cubev2/transformers_replace/models ${TRANSFORMERS_DIR}

Main Workflows

RoboTwin Evaluation

The detailed RoboTwin guide now lives in evaluation/RoboTwin/README.md.

Typical batch evaluation:

PRETRAINED_CKPT=zaleni/MagicBot-VGA-Robotwin \
QWEN3_VL_PRETRAINED_PATH=Qwen/Qwen3-VL-2B-Instruct \
QWEN3_VL_PROCESSOR_PATH=Qwen/Qwen3-VL-2B-Instruct \
COSMOS_TOKENIZER_PATH_OR_NAME=nvidia/Cosmos-Tokenizer-CI8x8 \
DISABLE_DA3_TEACHER_FOR_EVAL=true \
GPU_IDS=0,1 \
MAX_JOBS_PER_GPU=2 \
bash evaluation/RoboTwin/eval_randomized_50.sh

Real Piper Deployment

Piper deployment is documented in evaluation/Real_Piper/README.md.

Useful launch notes:

• CubeV2 Piper startup note
• MagicBot_R0 Piper startup reference

The Piper client supports sync inference, 7D real_piper state/action checks, and Enter-triggered return-to-init/restart when INIT_JOINT_POSITION and MANUAL_RESET=true are set.

Real Lift2 Deployment

Lift2 deployment docs live in evaluation/Real_Lift2/README.md.

The real-robot server code is shared by Lift2 and Piper where possible, while robot-side clients stay platform-specific.

Libero Evaluation

See evaluation/Libero/README.md.

Training Entrypoints

Common CubeV2 launch scripts:

• launch/cubev2/cubev2_pretrain.sh
• launch/cubev2/cubev2_finetune.sh
• launch/cubev2/cubev2_finetune_robotwin.sh
• launch/cubev2/cubev2_finetune_real_piper.sh
• launch/cubev2/cubev2_finetune_real_lift2.sh

MagicBot_R0 launch scripts:

• launch/magicbot_r0/magicbot_r0_pretrain.sh
• launch/magicbot_r0/magicbot_r0_finetune_robotwin.sh
• launch/magicbot_r0/magicbot_r0_finetune_real_piper.sh
• launch/magicbot_r0/magicbot_r0_finetune_real_lift2.sh

Norm-stat utilities for real-robot delta-action models are under launch/compute_norm/.

Tutorials

• Installation
• Finetune on LeRobot v2.1 dataset
• Finetune InternVLA-A1 with RoboTwin
• Pretrain InternVLA-A1 with InterData A1

External Assets

Depending on the model path, you may need local copies or Hugging Face repo ids for:

• Qwen3-VL backbone/processor, for example Qwen/Qwen3-VL-2B-Instruct
• Cosmos tokenizer, for example nvidia/Cosmos-Tokenizer-CI8x8
• DA3 teacher assets when training or when evaluation explicitly enables them
• MagicBot_R0 Wan/T5/VAE and ActionDiT/Future3D assets

For standard RoboTwin action evaluation with the released lightweight checkpoint, passing DISABLE_DA3_TEACHER_FOR_EVAL=true is recommended.

Acknowledgments

MagicBot-VGA started from the excellent InternVLA framework and has since been extended for CubeV2, MagicBot_R0, real-robot deployment, and multiple evaluation workflows.

We also thank these open-source projects:

• LeRobot
• RoboTwin
• Qwen3-VL
• NVIDIA Cosmos