A 1.7B Model Just Beat Its 6B and 8B Rivals: DexForce's UNITAS Treats Robot Manipulation as 3D Points, Not Pixels

A research team spanning DexForce Technology and several Chinese universities has released UNITAS, a robot manipulation model that abandons pixels for points. The paper, "UNITAS: A 3D-Native World Action Model for Embodied Manipulation," argues that most world action models built on pretrained video generators describe the world in images or latent visual features, while robot interaction happens in metric 3D space where pixel distance is not physical distance.

UNITAS instead places observations, actions and scene dynamics into a single shared 3D coordinate frame. An "action flow" writes human hands and robot grippers as 3D point trajectories, a "scene flow" describes how scene points move under those trajectories, and a physics-time trajectory tokenizer encodes each point path as a token anchored to its current 3D position. The backbone is a layer-coupled Mixture-of-Transformers that supports both direct action execution and action-conditioned scene prediction.

The results are reported by the authors and have not been independently reproduced. At 1.7 billion parameters, UNITAS claims an average success rate of 99.8% on LIBERO, 99.4% on the long-horizon LIBERO-Long subset, 89.1% on LIBERO-Plus, and 92.9% / 93.0% on RoboTwin 2.0 dual-arm manipulation in its Clean and Randomized settings — ahead of Fast-WAM at 6B parameters and Motus at 8B, and leading the π0.5, LingBot-VA and PointWorld baselines on several metrics.

On real hardware, the paper reports an average success rate of 85% across two tasks, with Organize Books at 90% and Pack the Ball at 80% over 20 trials each. In RoboTwin scene-motion prediction, mean trajectory error on moving points fell from 0.656 cm for PointWorld to 0.427 cm, a drop of about 35%, with 92.31% of predictions landing within 1 cm.

The ablations are the most interesting part. Removing the 3D positional encoding, action-flow supervision and action-flow history tokens collapsed overall success from 86.47% to 67.58%. Adding scene-flow supervision on top of the remaining design lifted success from 85.74% to 86.47% — evidence that learning how the world changes feeds back into better action decisions rather than being a decorative head.

Pretraining mixed nine data sources before task-specific fine-tuning. On LIBERO-Plus, mixed pretraining added 2.6 percentage points overall, and under perturbations of the robot's initial state it moved success from 84.8% to 94.2%. That is the paper's central claim about what a shared 3D representation buys: human-hand and gripper data become interchangeable training signal.

DexForce was founded in 2021 and builds generative world models and embodied AI infrastructure; its DexVerse engine generates 3D scenes from text and images for simulation. Code is available in the project's UNITAS repository on GitHub and the preprint is arXiv:2610.12099.

The broader significance is a bet that physical AI needs a native representation the way large language models needed text. Robot data is scarce and scattered across different bodies, cameras and sampling rates. If point trajectories in one metric space really do transfer across hands, grippers and simulators, then the value of every future robot dataset goes up — and that is a supply problem the field has not solved yet.