Figure Project Go-Big vs Egocentric-10K

Project Go-Big is a humanoid pretraining data-collection initiative that Figure announced in a company blog post on September 18, 2025, built for its Helix model with a Brookfield partnership. Egocentric-10K is a dataset that Build AI published on Hugging Face: 10,000 hours of head-mounted video of factory workers. One is an announcement of a collection effort and a single result; the other is a released corpus with a dataset card.

Both put egocentric human video at the centre of robot learning, but they are not like-for-like. Go-Big’s data is Figure’s own and no route for outside teams is described, while Egocentric-10K is obtainable behind a gate. Neither is a Firsthand product, and Firsthand has no relationship with Figure, Brookfield or Build AI. Firsthand data was not used in either. For the single-source explainers, see /guides/what-is-figure-project-go-big and /guides/what-is-egocentric-10k.

Figure states that Project Go-Big is accelerated by a partnership with Brookfield Asset Management, which it describes as covering over 100,000 residential units worldwide, about 500 million square feet of commercial office space and about 160 million square feet of logistics space. It says data collection has begun in Brookfield environments and will continue to scale.

The post reports one result. Figure trained Helix on 100% egocentric human video, collected passively in real Brookfield homes, with no robot demonstrations, and Helix now responds to commands such as “Walk to the kitchen table” by generating closed-loop navigation control from pixels. Figure says that, to its knowledge, this is the first time a humanoid has learned end-to-end from images and language to low-level SE(2) velocity commands using only human video. The post states no hour count, episode count, licence or benchmark number, so this page states none.

The card opens by calling Egocentric-10K “the largest egocentric dataset” and “the first dataset collected exclusively in real factories”. It reports 10,000 hours, 1.08 billion frames and 192,900 clips, with a median clip length of 180.0 seconds and a mean of 4.68 hours per worker. It also claims the dataset is “state-of-the-art in hand visibility and active manipulation density compared to previous in-the-wild egocentric datasets”. That is the publisher’s own description.

It is licensed under Apache 2.0 and gated: users must agree to share their contact information. The card text reviewed states no action labels, hand pose or language annotations, and reports no robot-learning result. It is a footage source with metadata, not a training recipe.

Go-Big’s video is egocentric and collected passively as people do everyday behaviors in real Brookfield homes. Figure argues the viewpoint is the point: a humanoid’s perspective and kinematics mirror a human’s, so skills can transfer from everyday video without costly robot demonstrations or staged environments. The capture hardware is not stated, so this page names none.

Egocentric-10K is its own capture, and the card is specific. The device is Build AI Gen 1, a monocular head-mounted camera recording 1080p at 30 fps with a 128° by 67° field of view and no audio. Footage is organised by factory and worker, from factory_001 to factory_085, in WebDataset shards of up to 1 GB. Each worker folder carries an intrinsics.json with an OpenCV fisheye calibration.

Go-Big is explicitly a humanoid pretraining effort. The human video is collected to train Helix, Figure’s vision-language-action model, and the reported result is a humanoid navigating from pixels after training on human video alone. Figure says a single Helix network now outputs both high-rate dexterous manipulation and navigation commands.

Egocentric-10K is a footage source not tied to a named robot. The card describes human workers with no robot embodiment, and the card text reviewed reports no robot-learning experiment. It sits in the human-video layer of a pretraining mixture, and a team still decides how to label it, how to mix it with robot data and how to evaluate transfer. See /guides/robot-pretraining-data.

Scale is measured differently. Egocentric-10K states data size: 10,000 hours, 1.08 billion frames and 192,900 clips. Go-Big states property access, which describes how many real spaces the partnership opens up, not how much video exists. A reader should not turn 100,000 residential units into an hour figure, or assume Go-Big is larger or smaller than 10,000 hours.

Access differs too. Go-Big is Figure’s own initiative for its own model, and the post describes no route for outside teams. Egocentric-10K is Apache 2.0 licensed and gated behind a contact-information agreement. An Apache 2.0 label is not a full licence chain: the card text reviewed does not describe consent or release terms for the workers filmed, so confirm those before building a product on it.

Egocentric-10K ships a JSON file beside each video with factory_id, worker_id, video_index, duration_sec, width, height, fps, size_bytes and codec, plus the per-worker intrinsics. It also links a 30,000-frame evaluation set, Egocentric-10K-Evaluation. No action, hand-pose or language annotations are stated.

Go-Big’s post says Helix produces low-level SE(2) velocity commands from pixels, but it does not describe how human video becomes a training signal. A buyer should not assume any particular method. For how another source handles human video with no action labels, see /compare/egocentric-10k-vs-gr00t-n1.

Neither source gives a buyer measured hand pose for its egocentric video, a consent chain for the people filmed that a funded product could rely on, or evidence that human video alone beats robot data. Go-Big runs no robot-data arm, and the Egocentric-10K card text reviewed reports no robot-learning result.

Beyond that, Go-Big does not state hours, episodes, licence or access, and Egocentric-10K does not cover anything outside factory work. Neither source says it uses the other’s data. The Go-Big post describes data Figure collects through its Brookfield partnership, and the Egocentric-10K card describes footage Build AI collected in factories.

Treat the two as evidence and a reference point, not a plan. Go-Big is evidence that egocentric human video from a humanoid-matching viewpoint is a viable pretraining source, from one company on one navigation task. Egocentric-10K is a concrete, obtainable scale reference for factory footage. Neither fixes how much data your robot needs, so start from your robot’s own coverage gaps. See /guides/why-dataset-diversity-matters-in-robot-learning and /compare/public-vs-custom-pretraining-datasets.

A fixed public corpus cannot be reshaped after release to your tasks, embodiment or environments. A custom collection starts with a written skill spec covering the scenes, tasks, objects and people a corpus is missing, proven with a fast, on-demand Spec Pilot before volume scales. Capture is head-worn, bimanual and whole-body for humanoid pretraining, with frame-accurate action segments and manipulation-grade 3D hand pose captured on the rig, ten synchronized streams on one hardware-triggered clock with a sub-2 ms ceiling, and metric depth alongside RGB, IMU and pose. Delivery is under a perpetual, buyer-owned commercial licence with a documented consent chain, in RLDS, LeRobot, WebDataset, HDF5, zarr and Rerun .rrd. See /custom-collection/humanoid-pretraining-data, /custom-collection/egocentric-data-collection-pilot and /custom-collection/pretraining-data-collection.