What Is UniAct (Universal Actions)?

UniAct is introduced in Universal Actions for Enhanced Embodied Foundation Models (arXiv 2501.10105, CVPR 2025), from authors at AIR at Tsinghua University, SenseTime Research, Peking University, Beijing University of Posts and Telecommunications and Shanghai AI Lab, with an official project page. The authors start from a familiar observation: foundation models gain from diverse internet-scale data, but embodied agents struggle to repeat that recipe because the action spaces of crowd-sourced robot datasets differ so much from one robot to the next.

UniAct is third-party research. Firsthand is not part of it or affiliated with its authors, and makes no claim that any model was trained on Firsthand data. What the site states is custom egocentric collection of the human side of a cross-embodiment mixture, fast and on demand, to a written spec: 3D hand pose at 21 joints per hand, calibrated stereo depth and 6-DoF IMUs at 200 Hz, delivered in RLDS, LeRobot format or WebDataset with a buyer-owned commercial license. Firsthand does not capture robot-embodiment action data. Whether UniAct uses human egocentric video, and how Firsthand’s signals would map to its universal action space, are Not stated in the sources read.

Per the abstract, many crowd-sourced embodied datasets exist, but their action spaces are heterogeneous because each robot has a distinct physical embodiment and control interface. That heterogeneity makes it hard to build embodied foundation models from cross-domain data, since the same label, such as an end-effector command, can mean different things on different machines.

UniAct is the authors’ answer: operate in one shared action space so that data from many robots can be used together instead of being siloed by control interface.

The authors describe a framework operating in a Universal Action Space. The learned universal actions capture generic atomic behaviors across diverse robots by exploiting their shared structural features. The project page calls them highly abstract, embodiment-agnostic intentions.

To act on a particular robot, a universal action is translated back into that robot’s heterogeneous actionable commands by adding embodiment-specific details. The authors say this makes fast adaptation to new robots simple and straightforward, and the project page demonstrates it with an interactive example on a WidowX robot.

The abstract reports that a 0.5B instantiation of UniAct outperforms state-of-the-art embodied foundation models that are 14 times larger, across real-world and simulation robots. The project page lists three settings: a real-world WidowX robot, a simulated Franka arm on LIBERO suites, and an AIRBOT tested under four different control interfaces for fast adaptation, with Octo and OpenVLA-7B as comparison models.

The project page gives per-setting tables, and results differ by platform and metric. Firsthand does not reproduce or compare those scores here, so read them at the source.

The site’s UniVLA entry covers a framework that derives task-centric latent actions from video. Its latent actions entry covers compressed representations inferred from what changes between frames, and its action-space alignment entry covers adapting human-derived actions with a smaller aligned robot dataset. UniAct is framed in its abstract around unifying heterogeneous robot action spaces into one universal space.

All of these are third-party ideas for the same broad problem of sharing skill across bodies, covered under cross-embodiment learning. Firsthand ranks none of them.

This entry draws on the arXiv abstract and the official project page. Training dataset composition and counts, whether human video is used, code availability and licence terms are Not stated in those sources, so none appear here.

When a program needs real human egocentric demonstrations of tasks, environments or objects that public data does not cover, specified in writing, as the human side of a pretraining mixture. Firsthand does not build or train models.

Confirm the spec on a Spec Pilot, a small fast first batch. Where coverage is deep, the median from signed spec to first validated batch is 48 hours; where it is thin, three to four weeks. Billing is per validated hour.

What is UniAct?

A published embodied foundation modeling framework (arXiv 2501.10105, CVPR 2025) that operates in a learned Universal Action Space shared across different robots.

What are universal actions?

Per the paper, learned representations of generic atomic behaviors shared across diverse robots through their common structure. They are translated back into each robot’s own commands by adding embodiment-specific details.

Why does UniAct need a universal action space?

The abstract says crowd-sourced embodied datasets have heterogeneous action spaces because robots differ in embodiment and control interface, which makes cross-domain training hard.

How well did UniAct perform?

The authors report that a 0.5B instantiation outperforms embodied foundation models 14 times larger across real-world and simulation robots. Scores vary by setting, so see the project page tables. Firsthand makes no comparison from them.

Does Firsthand data train UniAct?

No claim is made. Firsthand is not affiliated with its authors, and does not state that any model was trained on Firsthand data.

Can Firsthand collect egocentric data for a cross-embodiment mixture?

The site states custom egocentric collection to a written spec, fast and on demand, with 21-joint 3D hand pose, calibrated stereo depth and 200 Hz IMUs, confirmed on a Spec Pilot. It does not capture robot-embodiment action data.