Open Source Robotics

A curated collection of high-quality open source robotics projects, tools, and software to propel the robotics community forward.

ABC-130k

Largest open-source bimanual robot teleoperation dataset: 130,919 episodes

License

Apache-2.0

Overview

ABC-130k is the largest open-source robot teleoperation dataset. It contains 130,919 bimanual manipulation trajectories—3,598 hours in total—collected on two-arm YAM stations, each equipped with two 6-DOF arms, parallel-jaw grippers, and three cameras (one overhead and two wrist-mounted). 43,090 of the episodes carry subtask annotations, stored separately from the episode data. The dataset is released under Apache-2.0 in MCAP format, with joint states, gripper positions, and synchronized video streams, alongside a companion research paper and training/deployment code. Created by Amazon FAR.

ABC-130k targets the central bottleneck in generalist manipulation—the scale and diversity of real robot data—and does so for the under-served bimanual setting. Every trajectory is a human teleoperation demonstration on a standardized two-arm YAM station, captured as MCAP with full joint/gripper state and three synchronized camera views. A large annotated subset adds subtask labels for finer-grained learning and evaluation. At 130,919 episodes and 3,598 hours it is, to date, the largest openly licensed teleoperation corpus, and it ships with code for training and deploying policies on the data.

Why it matters

Generalist manipulation policies are bottlenecked by the scale and diversity of real robot data, and most large open datasets are single-arm. ABC-130k is both foundation-scale and bimanual, covering the two-arm coordination that the field most lacks, and gives the community an openly licensed resource for training and benchmarking bimanual manipulation. Roughly 32,000 downloads per month signal strong, immediate adoption.

Problem it solves

Open robot-learning data is largely single-arm and limited in scale; there was no foundation-scale, openly licensed bimanual teleoperation dataset for training and benchmarking two-arm manipulation.

Key takeaways

  • Largest open teleoperation dataset to date: 130,919 episodes, 3,598 hours, ~22 TB
  • Bimanual two-arm YAM stations: two 6-DOF arms, parallel grippers, three cameras
  • 43,090 episodes with subtask annotations stored separately from episode data
  • MCAP format with joint/gripper state and synchronized video, plus paper and training/deployment code
  • Apache-2.0 licensed; ~32k downloads/month

Who it is for

  • Bimanual manipulation researchers
  • Robot foundation-model trainers
  • Imitation-learning teams
  • Benchmark and evaluation builders

Limitations

  • ~22 TB total—significant storage and bandwidth requirements
  • Collected on a single hardware platform (two-arm YAM stations)
  • Teleoperation demonstrations, not autonomous or optimal trajectories
  • Subtask annotations cover a subset (43,090 of 130,919 episodes)

This website requires JavaScript to function properly. Please enable JavaScript to explore our collection of open source robotics projects.