amazon-far/TWIST2

[arXiv 2025] TWIST2: Scalable, Portable, and Holistic Humanoid Data Collection System

What it solves

TWIST2 addresses the difficulty of collecting high-quality, scalable, and portable data for humanoid robots. It provides a system to bridge the gap between human motion and robot execution, allowing for efficient teleoperation and data collection in both simulation and real-world environments.

How it works

The system separates high-level motion control from low-level robot execution.

  1. High-Level Control: Human motions are captured and streamed via a PICO VR headset (online teleoperation) or from offline motion datasets.
  2. Motion Retargeting: It uses a General Motion Retargeting (GMR) system to map human movements to the robot's morphology.
  3. Low-Level Control: A trained RL-based general motion tracker (student policy) executes the retargeted motions. This policy is exported as an ONNX model for efficient deployment.
  4. Communication: A Redis server is used to handle the communication between the high-level motion streaming and the low-level controller.

Who it’s for

Researchers and engineers working on humanoid robotics, specifically those focusing on imitation learning, sim-to-real transfer, and large-scale humanoid data collection.

Highlights

  • Holistic Data Collection: Supports both simulation (sim2sim) and physical robot (sim2real) deployment.
  • VR Teleoperation: Integrates with PICO VR headsets for real-time human-to-robot control.
  • Modular Architecture: Decouples high-level motion streaming from low-level RL policies.
  • Hardware Support: Specifically designed for and tested with the Unitree G1 humanoid robot.

Related

  • Project
    locusrobotics/fuse

    A ROS framework for live sensor fusion on robots using nonlinear least squares optimization to handle state estimation, localization, and calibration.

  • Project
    KINGSTON-115/llm-pid-tuner

    An LLM-assisted utility for PID tuning that analyzes system responses from hardware or simulations to automatically suggest and apply optimal control parameters.

  • Project
    UniversalRobots/Universal_Robots_ROS_Driver

    A ROS driver for Universal Robots CB3 and e-Series manipulators that provides a stable interface for real-time control, calibration, and integration with ROS-control.

  • Project
    icey1287/SuperMew

    SuperMew is a self‑hosted, audit‑first RAG/agent platform. It persists threads, runs, events and document versions, offers hybrid Milvus+BM25 retrieval, supports HITL checkpoints, versioned skills/tools, a frozen model control plane, and a built‑in evaluation framework. The stack uses FastAPI, PostgreSQL, Redis, Milvus, and a Vue 3 frontend, and can be run locally with Docker‑Compose or deployed in production with systemd/K8s.