Neocambrian AI launches Mitts — tactile gloves for high-precision robot touch data

TL;DR

Neocambrian AI (“Building the internet of human action”) launches Mitts — wearable tactile gloves instrumented with 450+ sensing points across the palm and fingers, each detecting as little as 5 g of force at 30 Hz. The stated purpose is to capture high-precision whole-hand touch data from humans at scale as training supervision for robot manipulation policies, synchronized with existing phone or camera rigs (iPhone, BitRobot Network’s RoboCap). The company says the gloves are already deployed collecting data across industrial and commercial environments and is soliciting partners.

Key claims

• Whole-hand coverage with 450+ sensing points across palm and fingers at ~5 g sensitivity and 30 Hz sampling [tweet thread, part 2/3].
• Hardware synchronization with iPhone and with BitRobot Network’s RoboCap capture rig, so tactile is time-aligned with egocentric/third-person RGB streams [tweet thread, part 2/3].
• Positioned explicitly against vision-only recipes: “Vision alone isn’t enough” [tweet thread, part 1/3].
• Already in use collecting tactile data across industrial and commercial environments, with partnerships open [tweet thread, part 3/3].

Method

The product is a soft wearable glove with distributed pressure/contact sensors; no engineering disclosure beyond the headline numbers (450+ points, 5 g minimum detectable force, 30 Hz) and the sync surface (iPhone, RoboCap). The implied data pipeline is: human wears glove while performing a task in a real environment → tactile stream recorded in sync with camera video → used as training supervision for a downstream robot policy (hand or gripper). No sensor abstraction (per-taxel force vector, binary contact, pressure heatmap, elastomer displacement) is disclosed; no sim-to-real story is offered; no accompanying model, policy, or dataset release is announced.

Results

Announcement only — no quantitative results, benchmarks, or policy evaluations shared. The reply thread includes third-party reactions but no deployment numbers.

Why it’s interesting

Mitts stakes out a data-collection-hardware-first position on the long-running “where does touch enter the pipeline” debate the Tactile sensing for manipulation cluster tracks: not at the policy (Tactile-Reactive Dexterous Hand: High-Frequency Physical Interaction T-Rex), not at the simulator (Tactile Genesis: Exploring Tactile Sensors at Scale for Learning Dexterous Tasks Tactile Genesis), not inferred from vision (PressureVision++: Estimating Fingertip Pressure from Diverse RGB Images PressureVision++, Being-H0.8: A Latent Tactile World-Action Model at Scale Being-H0.8), but captured at human-demonstration scale via a wearable the demonstrator wears in the field. The 450-point whole-hand coverage directly operationalizes Tactile Genesis’s “placement dominates sensor type / whole-hand coverage beats fingertip-only” design finding — more so than fingertip-focused rigs like Aero Hand Open (Aero Hand Open: low-cost, open-source, lightweight, anthropomorphic dexterous hand) or ORCA (ORCA: An Open-Source, Reliable, Cost-Effective, Anthropomorphic Robotic Hand for Uninterrupted Dexterous Task Learning). On the collection-substrate axis it is the glove sibling to ACE-Data-0’s full-palm tactile grids (ACE-Data-0: Human-Centric Ambient Capture as Embodied Data Engine) and to Sharpa’s bimanual F/T Origami Dataset (Sharpa releases the Origami Dataset for the IROS 2026 Robotic Origami Challenge), and plugs into the same RoboCap capture surface as Pablo Vela Gomez’s SLAM-via-cuVSLAM-on-RoboCap work (SLAM on Robocap via cuVSLAM, packaged with rattler-build (Claude Code build report)). The open question the announcement does not answer is which sensor abstraction the recorded stream exposes to downstream training — the design axis that distinguishes Tacmap (Tacmap: Bridging the Tactile Sim-to-Real Gap via Geometry-Consistent Penetration Depth Map) from Tactile Genesis, and that Open-X-Tactile (Open-X-Tactile — community effort to build the world's largest heterogeneous tactile manipulation dataset) is explicitly trying to standardize across contributing labs.

See also

• Tactile sensing for manipulation — wearable glove collecting data at the frontier of the “coverage over fingertip-only” directive
• Human-to-Robot Retargeting — human-worn capture device whose stream feeds into human-video → robot retargeting pipelines
• ACE-Data-0: Human-Centric Ambient Capture as Embodied Data Engine — ACE-Data-0 ships full-palm tactile grids synced to multi-view human video; the closest filed sibling on the glove-based substrate
• Sharpa releases the Origami Dataset for the IROS 2026 Robotic Origami Challenge — Sharpa Origami Dataset bundles bimanual F/T + vision synced at 30 FPS, same frame rate target as Mitts
• Open-X-Tactile — community effort to build the world's largest heterogeneous tactile manipulation dataset — Open-X-Tactile multi-lab coordination effort that any Mitts-captured corpus would plausibly contribute to
• SLAM on Robocap via cuVSLAM, packaged with rattler-build (Claude Code build report) — BitRobot RoboCap sync target is the same rig used here for cuVSLAM experiments
• Tactile Genesis: Exploring Tactile Sensors at Scale for Learning Dexterous Tasks — simulator-side ablation that Mitts’s 450-point whole-hand coverage operationalizes on the hardware side