RewardAI OM-1 + Omnibody Hand launch — zero-shot cross-embodiment dexterous manipulation from human data (DexCap successor)

TL;DR

RewardAI (@rewardai_) launched OM-1, a dexterous-manipulation model trained to capture “subconscious human physical intelligence” — nudging, sliding, twisting, power grasping, whole-body coordination — from human demonstration data, framed as the successor to the group’s earlier DexCap system at Stanford. Two claims stake out the launch: (a) zero-shot cross-embodiment generalization across small arms, industrial arms, and humanoids paired with their Omnibody Hand (7-DoF anthropomorphic design blending thumb-index precision and coordinated-finger power grasps), and (b) real-time execution at deployment on long-horizon tasks (quadmanual phone packaging, bartending, laundry folding, Ethernet-cable unplugging), with all showcase videos claimed to be unspeedd. Emergent behaviors are also claimed — bimanual mistake compensation, retry, adaptation to adversarial perturbations, task abandonment on drastic environment changes. Launch-thread only; no paper, checkpoint, training corpus, per-task success rates, backbone / action-head architecture, or head-to-head benchmarks disclosed.

Key claims

OM-1 captures human physical-manipulation intelligence directly from humans (nudging, sliding, twisting, power grasping, whole-body coordination), building on the group’s prior DexCap system at Stanford [tweet 1, tweet-body].
OM-1 zero-shot generalizes across embodiments — small arms, industrial arms, humanoids — with the 7-DoF Omnibody Hand, without robot-specific finetuning [tweet 2].
The Omnibody Hand is a compact 7-DoF anthropomorphic design targeting useful contact points, in-hand reorientation, and smooth transitions between precision (thumb-index) and power (coordinated-finger) grasps [tweet 3].
Showcase tasks include quadmanual phone packaging, bartending, and laundry folding, presented as <30 s long-horizon jobs with real-time (no-speedup) execution [tweet 4].
Emergent capabilities claimed from the same policy: bimanual mistake compensation between arms, retry, adaptation to adversarial perturbations, and stopping when environment changes exceed some threshold [tweet 8].
Deceptively-hard fine-grained task tested: unplugging an Ethernet cable (locking latch must be pressed precisely before release), reliably solved [tweet 5].
Whole-body humanoid manipulation and navigation are teased as running on OM-1 as well, with details deferred [tweet 7].

Method

Launch thread only — no method disclosed beyond the framing that OM-1 is trained on human demonstration data captured via a DexCap-style egocentric mocap pipeline (the linked project page describes DexCap’s original recipe: chest-mounted RGB-D + SLAM cameras, EMF gloves for finger motion, mocap-glove hand tracking retargeted to a LEAP-style multi-fingered robot hand, Diffusion Policy trained on retargeted point-cloud observations, optional human-in-the-loop residual correction and full teleoperation for fine-tuning). The Omnibody Hand replaces LEAP as the target embodiment: 7 DoF, compact form factor, coordinated multi-finger mechanism, targeting both fingertip-precision and whole-hand power grasps. The embedded DexCap link is the strongest signal to the underlying data-collection substrate; the training recipe (backbone, action head, loss, corpus scale) is not disclosed in the launch thread.

Results

Videos in the thread show:

Small-arm, industrial-arm, and humanoid embodiments running the same OM-1 + Omnibody Hand pairing without per-embodiment finetuning [tweet 2].
Quadmanual phone packaging, bartending, and laundry folding in real time (no speedup) [tweet 4].
Ethernet-cable unplug (locking-latch fine-grained manipulation) [tweet 5].
Humanoid whole-body manipulation and navigation [tweet 7].
Bimanual arm-to-arm mistake compensation, retry, adversarial-perturbation recovery, task abandonment [tweet 8].

No quantitative results (per-task success rate, N-of-rollouts, comparison against π0.5 / GR00T / other VLAs, teleop-data-hour count, per-embodiment adaptation cost) are reported in the thread. Independent third-party evaluation (e.g. via Inspect Robots — An Open-Source Evaluation Framework for Physical AI (Robocurve) on YAM-arm hardware, or RoboArena: Distributed Real-World Evaluation of Generalist Robot Policies) does not exist at filing time.

Why it’s interesting

OM-1 stakes out a dexterous-hand-first variant of the “make the collection tool embodiment-invariant” recipe already contested by Solving Dexterity: A Full-Stack Approach (mimic hand M1 + wearable U1) (mimic M1 hand + U1 wearable), Xiaomi-Robotics-1 (XR-1) — Scaling VLA Foundation Models with 100K Hours of Embodiment-Free UMI Pre-training (XR-1 100K-hour UMI corpus with parallel grippers), and TwinDEX: A Scalable Path from Robot-Free Data Collection to Real-World Dexterous Manipulation (TwinDEX bimanual three-finger nine-DoF wearable) — but with the demonstrator side pushed back to bare-hands human mocap via the DexCap glove-based capture rig rather than a mechanically-locked wearable that constrains the demonstrator into the target hand’s motion space. The 7-DoF Omnibody design also positions OM-1 between the low-DoF parallel-gripper UMI lineage and the 15+-DoF fully-anthropomorphic hand lineage (Aero Hand Open: low-cost, open-source, lightweight, anthropomorphic dexterous hand Aero Hand, mimic M1) — an interesting mid-DoF point on the same axis the Human-to-Robot Retargeting concept tracks. The zero-shot cross-embodiment claim (small arm + industrial arm + humanoid all driven by one policy) puts OM-1 in the same “one policy across many arms and one hand” position as Cross-Embodiment Robot Manipulation via a Unified Hand Action Space UHAS (unified sphere-fingertip action space) and RealDex1 (How we actually test a robot hand RLWRLD’s hand-agnostic VLA foundation model), but with the invariance carried by a single-hand-hardware + policy pairing rather than an embodiment-invariant action space. The launch is a press-release-shape artifact: none of the load-bearing quantitative claims (task success, generalization gap, per-embodiment adaptation cost, teleop-hour count) have numbers, and the DexCap link is the only technical anchor.