Action Intelligence launches Chorus: robot training data grows 40x

Action Intelligence (行元智能) announced Chorus ("众行" in Chinese) on October 27 — a cross-embodiment data engine that turns video of human actions into pretraining data for robots of different physical designs. The claimed effect: a 40x increase in the training dataset. Video sources are collected from five types of wearable and handheld devices, and the main goal is to boost the reuse rate of human data.

Why human video became robotics' new currency

Language models had the internet. Robots don't. Every hour of real-world object manipulation has to be either recorded via teleoperation — where an operator wearing a headset or holding controllers drives a manipulator arm — or generated in simulation, which transfers poorly to the physical world. The first path is expensive and slow: it requires robots, test sites, and trained operators. The second suffers from the sim-to-real gap.

Human video works differently. People already pick up cups, open drawers, and fold clothes every day. Capture that on cheap wearable cameras, and data accumulates orders of magnitude faster. But there's a catch: a human hand has five fingers and its own kinematics, while a robot might have a two-finger gripper, an anthropomorphic hand, or a manipulator with a different number of degrees of freedom. That's the embodiment gap. Chorus closes it at the data level, converting source video into a form usable for a specific platform.

Five device types feeding one pipeline

The strength of the announcement is that Chorus isn't tied to one capture method. The company lists the following sources:

• Smart glasses. Provide a first-person view closest to what a robot's "head" would see.
• Chest-mounted camera. Captures the work area in front of the body and the hands in frame without restricting movement.
• Wrist-mounted camera. Shows object grasping in close-up, useful for training precise contact actions.
• UMI devices. Handheld grippers with a camera, based on the Universal Manipulation Interface. Human motion is recorded in a form easier to transfer to a manipulator arm.
• Smart head band. A lightweight wearable sensor for long filming sessions.

Each device distorts the picture in its own way. Glasses capture the whole scene but hands can get occluded. A wrist camera shows contact precisely but loses context. A single engine that ingests such a heterogeneous stream spares customers from building a separate pipeline for each sensor.

What the 40x increase actually promises

The figure of 40 in the release headline refers to the volume of training data available for pretraining after human videos are processed through Chorus. The logic is straightforward: if the same recording can be used for multiple robot types instead of just the platform it was filmed on, the total usable volume multiplies. Pretraining robotics foundation models has long been bottlenecked precisely by diversity and scale — without tens of thousands of hours across different tasks, scenes, and objects, models generalize poorly.

The key question that demands caution: how well does the converted data actually perform on real robots. This is exactly where cross-embodiment approaches are usually tested — by comparing task success rates for models trained with and without added human data. There are no independent benchmarks for Chorus yet, so it will have to be judged by future demonstrations and open benchmarks.

Who stands to benefit in practice

For humanoid and manipulator developers, Chorus promises to reshape the cost structure: part of the teleoperation budget could be replaced by cheap human filming. Teams that already have their own fleets of wearable cameras get a chance to reuse accumulated archives. Integrators and factories where operators perform routine assembly and warehouse tasks gain the ability to record real staff work and turn it into training material.

Three things are worth watching: exactly which robots are supported, what format the engine outputs data in, and whether Action Intelligence will publish comparisons against baseline methods. The answers will determine whether Chorus becomes a working infrastructure layer or remains a polished presentation.