• 100 hours of Human Archive data.
• 100 hours of Mecka data from EgoVerse, randomly sampled.
• 100 hours of microagi data from EgoVerse, randomly sampled.

We used 100 hours of data from our OTS catalog. We then fine-tuned each pretrained model, plus a fourth copy with no human-video pretraining, on the same robot demonstrations. Finally we tested all four models on a real Franka arm with a Wuji dexterous hand: 200 trials across four tasks.

Human Archive data had the highest task success on objects the robot saw during fine-tuning, and we saw our largest win on objects it had never seen during fine-tuning.

Results

All results

Human Archive data is competitive with the best egocentric datasets available today. In this test it led on both familiar and new objects in task success.

Why we think this happened

Human Archive is already working alongside leading AI labs to scale human video and sensor data. Our data is SOTA because it is:

• Extremely diverse. We collect across tens of thousands of businesses and unique individuals, across every type of commercial business and households. We currently have data-operations partnerships with some of the largest businesses in Asia.
• Strong scene quality and behavioral coverage. Task complexity, dexterous interactions, hands in frame, natural head motion, occlusions, and variation in how each task is performed.
• SOTA 3D action labels and dense text labels. Our hand pose annotation is millimeter-level in the wild (see HA-SLAM and HA-Ego500).
• Collection rigs designed specifically for egocentric data. Multi-camera with <0.5 px reprojection error, 60 fps, egocentric global shutter.

Setup

Policy

VITRAVision-language-action model from Microsoft Research Asia and Tsinghua University (Li, Deng, Liang et al., 2025, “Scalable Vision-Language-Action Model Pretraining for Robotic Manipulation with Real-Life Human Activity Videos”, arXiv:2510.21571)

Robot

Franka Research 3 (FR3) arm with a Wuji Hand 220-DoF dexterous hand

Camera

One fixed Luxonis OAK-D WRGB-D camera; the model used the 1280×720 RGB stream

Pretraining data

~100 hours per datasetEgocentric human video with 3D hand tracking, SLAM camera poses and text annotations

Fine-tuning data

Teleoperated robot demonstrationsCollected with an Apple Vision Pro (hand tracking retargeted to the Wuji hand)

Tasks and demonstrations

Held-out demos are used for checkpoint selection.

Objects

Seen in training

New objects

Model inputs and outputs

• Inputs: one RGB frame from the fixed camera, cropped and resized; the language instruction; and the robot's current state (wrist pose and finger positions, expressed in the same hand format as the human data).
• Outputs: a chunk of 16 future actions at 15 Hz (about 1.07 s). Each action is a wrist movement (translation and rotation) plus finger joint angles: inverse kinematics turns the wrist movement into arm targets, and the finger joint angles drive the 20 joints of the Wuji hand.

Because robot actions use the same hand format as the human videos, what the model learns from people can carry over to the robot.

How the robot runs the model

• The control loop runs at 15 Hz.
• For each prediction, the robot executes the first 8 of the 16 actions (0.53 s) exactly as predicted. It then holds still while the next chunk is computed from a new camera frame.
• The model runs on an NVIDIA RTX 5090 over Wi-Fi. A round trip takes a median of 0.43 s (0.32 s of that is spent on the server).
• We sample 10 diffusion (DDIM) steps with classifier-free guidance of 1.5, in bf16.

We deliberately did not use real-time chunking (RTC) or any other asynchronous execution, so the robot pauses briefly between chunks, and you can see it in the videos. The goal of this study was to make the models work and evaluate them honestly under identical, simple conditions, not to make them fast. Every model ran under exactly the same execution scheme.

Training

Pretraining

One run per dataset, with identical code and configs; only the data differs. We followed VITRA's stage-1 recipe, scaled to 100 hours:

• Start from Google's PaliGemma-2 3B (paligemma2-3b-mix-224) and a freshly initialized diffusion-transformer action head (DiT-B).
• 20,000 steps at a global batch of 512, about one pass over 100 hours, on 4× NVIDIA H100.
• Learning rate 1e-5 for the vision-language model and 1e-4 for the action head. The vision encoder stays frozen, and the language model stays frozen for the first 1,250 steps.
• Left/right flip augmentation, as in VITRA.
• Normalization statistics computed per dataset.

Fine-tuning

Same recipe for all the tasks; only the starting weights differ.

• AdamW, learning rate 1e-5, weight decay 0.1, global batch 256.
• Vision encoder frozen, no flip augmentation.
• Lid and bottle (one model, 40 demos): 1,000 steps. Rubik's cube: 1,000 steps. Dunk: 1,500 steps.
• The no-pretraining model got more steps because it learns more slowly: 1,500 instead of 1,000, and 2,000 instead of 1,500 for dunk.
• We validated on the held-out demos every 50 steps (every 100 for dunk) and deployed the checkpoint with the lowest wrist-trajectory error.

Testing

• 200 trials: 4 models × 50 trials. Each model got 15 lid trials, 15 bottle trials, 10 dunk trials and 10 Rubik's cube trials, for 25 trials on seen objects and 25 on new objects.
• Interleaved: we tested in rounds. Each round ran every model once, back to back, so lighting and robot condition stayed matched across models.
• One-minute limit: a trial ended at success or after 1 minute.
• Graded live by the operator.

Grading

Each trial gets the highest stage it reached. Success means reaching the last stage.

Spatial generalization

Every task was built for spatial generalization: we moved the objects around the table both while collecting demonstrations and during testing. The lid is one example. When we moved it to a new spot on the table, the Human Archive model still found it and closed it.