UniACT Dataset Deep Dive: How AMAP Unified 6M+ Trajectories for Cross-Embodiment VLA Training

On March 31, 2026, AMAP CV Lab (AutoNavi) released the UniACT dataset alongside the ABot-M0 model — a massive embodied AI data resource integrating 6 major open-source datasets, covering 20+ robot embodiments with over 6 million trajectories.

What Is UniACT

UniACT is not a newly collected dataset from scratch, but rather a data standardization layer. It takes 6 heterogeneous datasets — OXE, OXE-AugE, AgiBot-Beta, RoboCoin, RoboMind, and Galaxea — and processes them into a consistent training corpus.

The core challenge: different datasets use different action spaces, sensor configurations, and robot morphologies. Open X-Embodiment uses absolute end-effector poses, AgiBot-Beta uses joint angles, and Galaxea involves dual-arm coordination — fitting them into the same training batch requires a unified representation.

Technical Core: Unified Action Representation

UniACT's solution:

• Action Unification: All data is converted to end-effector delta pose + rotation vector, eliminating differences between absolute pose and joint space
• Dual-Arm Padding: Single-arm datasets are automatically padded to a dual-arm format with zeroed missing-arm actions
• Standardized Pipeline: The official GitHub repo provides complete data construction scripts — one command to reproduce

The direct beneficiary is VLA (Vision-Language-Action) models. ABot-M0, pretrained on UniACT, demonstrates strong generalization across multiple manipulation benchmarks.

Scale & Coverage

• 6M+ trajectories, 9,500+ hours of data
• 20+ robot embodiments: from single-arm WidowX to dual-arm humanoid robots
• 6 source datasets: covering grasping, manipulation, long-horizon tasks
• Apache 2.0 license: commercially usable

Name Confusion Alert

Do not confuse UniACT (AMAP) with UniAct by Tsinghua (2025) — the former is a dataset released by AMAP CV Lab, the latter is a universal action framework proposed by Tsinghua/SenseTime. Nearly identical names, completely different things.

How to Access

UniACT does not offer a one-click download — you need to clone the official repo, download the 6 source datasets, then run the standardization scripts to build it yourself. The repo also provides pretrained ABot-M0 weights for direct inference or fine-tuning.

GitHub: github.com/amap-cvlab/ABot-Manipulation

Project Page: amap-cvlab.github.io/ABot-Manipulation

Paper: arxiv.org/abs/2602.11236

Disclaimer: All information above is sourced from publicly available materials (GitHub, arXiv, project pages). This post is for informational and technical analysis purposes only and does not constitute commercial advice. Dataset copyrights belong to their respective owners/authors. Corrections are welcome via the comments section below.