Study A
Does measured joint data train a better policy?
We train the same policy twice, once on measured joint angles and once on pose estimated from video, then compare success rates. The comparison gets cited a lot and has never been published.
Research
The two experiments
the field hasn't run.
Both studies publish with the dataset. Every claim on the landing page links to its paper further down this page.
Study B
Can a robot hand use force recorded from a human?
We record force from a human hand, map it onto a robot hand that can feel, and run the whole path from recording to trained policy.
The sources, in full
01 / 05
Source · ablation study
T-Rex: touch beats scale on contact-heavy tasks
34 authors · Stanford / Berkeley / NVIDIA / Meta affiliations · 2026
A 100-hour force-annotated demonstration corpus outperformed a 20,854-hour egocentric-video pretraining corpus on contact-heavy manipulation. The controlled ablation isolates the force channel itself: adding it moves success from 42% to 65%.
Open the paper →65%success · 100 hrs w/ force
35%success · 20,854 hrs video
+23 ptsforce channel, isolated
Source · fragile manipulation
Feel the Force: force-aware policies on fragile objects
2025 to 26 force-aware imitation literature
Policies that track a predicted force target succeed on fragile and deformable objects where force-blind baselines fail outright. The gap is 87% versus 0%. Without pressure feedback the policy cannot modulate grip at all.
Open the paper →87%force-aware success
0%force-blind success
Source · hand-pose benchmarks
Vision-based hand pose collapses under occlusion
Transformer hand-reconstruction benchmarks (HaMeR-class models)
On standard benchmarks, state-of-the-art hand-pose estimation scores roughly twice as well on visible joints as on occluded ones. During a grasp the fingertips are behind the object at exactly the frames that decide success. Measured joint sensing has no such failure mode.
Open the paper →~52 to 61PCK · visible joints
~23 to 27PCK · occluded joints
Source · public-corpus survey
Force data is ~200× scarcer than video
Survey of public manipulation corpora · 2026
Comparing public egocentric-video hours against the largest public force-sensing corpus yields a verified ratio of roughly 208:1. Video can be scraped from the internet; calibrated pressure at the moment of contact can only be recorded at the source. It does not exist anywhere else.
Open the paper →208:1video hrs : force hrs
0scrapeable force datasets
Source · Meta FAIR
Meta FAIR on the limits of video
Open-sourced force-sensing glove research · Meta FAIR
Meta’s own glove-sensing research states it directly: “video alone cannot capture the rich contact signals critical for mastering manipulation.” They open-sourced the sensor designs and shipped no product.
Open the paper →12force sites validated
0products shipped
Side by side
How robots learn manipulation today.
The force column is where the table thins out.
| Data source | Finger position | Grip force | During contact (fingers hidden) |
Cost to scale |
|---|---|---|---|---|
| Internet & egocentric video | Guessed from pixels | None | Accuracy halves | Cheap, scrapable |
| Vision hand-tracking | Estimated | None | Fails at the grasp | Cheap |
| Robot teleoperation | Robot joints | Rare, uncalibrated | Yes | A robot per operator |
| Simulation | Perfect but synthetic | No validated contact physics | Yes | Near-free |
| Digitus glove | Measured · 26 DoF @ 120 Hz | Calibrated · 0.1 to 100 N | Unaffected | Human-speed, no robot |
The episode format
The format the glove exports.
Force and position land on one clock and export as LeRobot-compatible episodes. The sample below comes from the same grasp model as the animations on this site and carries a simulated watermark.
EPISODE FORMAT · TARGET · LEROBOT-COMPATIBLE
observation.joints float32[26] @ 120 Hz observation.force float32[3×3] normal + shear timestamp int64 one clock sync_offset_ms float32 <10 verifiedDownload a sample episode (simulated) ↓
The dataset
The dataset both studies publish with.
Leave an email and we will send the preprint and the download once they go live.
Your email draft is open. Press send in your mail app and you are on the list.