amazon-far/abc

ABC: Scalable Behavior Cloning with Open Data, Training, and Evaluation

What it solves

ABC provides a scalable framework for behavior cloning, enabling robots to learn complex tasks from open data. It addresses the challenge of training robust policies that can generalize across different tasks and environments, providing both the model architectures and the necessary tools for training, evaluation, and real-world deployment.

How it works

The project implements two primary model architectures:

  • ABC-DiT: A Diffusion Transformer (DiT) policy that uses a vision backbone (defaulting to DINOv3 or CLIP) to process visual inputs and generate robot actions.
  • ABC-VLA: A Vision-Language-Action (VLA) model based on a diffusion-only port of Gemma 3 4B, utilizing SigLIP for vision and a DiT head for action generation.

The system includes a self-contained MuJoCo-based simulator for task evaluation, a data conversion pipeline for processing MCAP files, and a deployment stack for real-robot inference and teleoperation.

Who it’s for

Robotics researchers and engineers who need to train and deploy behavior cloning policies using large-scale datasets, specifically those interested in high-capacity models like Diffusion Transformers and VLAs.

Highlights

  • Dual Architecture Support: Offers both a specialized DiT policy and a large-scale VLA model based on Gemma 3.
  • Integrated Simulation: Includes a comprehensive MuJoCo simulator with a task catalogue and randomization tools for rigorous evaluation.
  • Real-Robot Ready: Provides a full deployment stack including RTC, teleoperation, and DAgger checkpoints for physical hardware.
  • Open Data Pipeline: Supports the ABC-130k dataset and provides utilities to convert raw data into training-ready formats.

Related

  • Project
    InternRobotics/VLAC

    A Vision-Language-Action-Critic model that provides dense reward feedback and trajectory quality screening for real-world robotic reinforcement learning.

  • Project
    hzxie/DynamicVLA

    A Vision-Language-Action (VLA) model and dataset (DOM) designed for robots to perform dynamic object manipulation in simulated environments.

  • Project
    Lei-Kun/RL-100

    A real-world reinforcement learning framework for robotic manipulation that enables a unified pipeline from imitation learning to offline/online RL and one-step policy distillation.

  • Project
    dexmal/opendm

    OpenDM is a Vision-Language-Action (VLA) model and framework for open-world robot control, enabling robots to follow open-ended instructions and handle complex, long-horizon tasks.

  • Project
    X-Square-Robot/wall-x

    An open-source framework for embodied foundation models that enables robots to perform complex physical manipulations by bridging vision-language understanding with action prediction.