DexForce's UNITAS Gives Robot Actions a Shared 3D World Model

How UNITAS Connects Motion With Its Consequences

A video can show a gripper moving toward a cup without directly specifying the distance between them. Camera placement changes the apparent motion, and equal movements across an image need not represent equal movements in a room. UNITAS makes physical coordinates part of the learning interface, rather than asking the model to infer all those relationships implicitly from pictures.

The preprint describes two related predictions. Action flow tracks points on a robot gripper or human hand through three-dimensional space. Scene flow tracks how points in the surrounding scene move in response. A third output supplies the executable robot commands. This separates a proposed physical movement, its expected environmental effect and the controls needed to carry it out.

That arrangement also helps combine training material from different sources. The researchers describe nine sources spanning simulation, real robots and human-hand demonstrations. Their method uses a common point-trajectory representation while accounting for elapsed time, so recordings sampled at different rates do not automatically become equivalent motion sequences. This is a method for learning across heterogeneous data, not evidence that simulation has eliminated the need for real-world demonstrations.

Under the Hood: A Shared Metric Frame

UNITAS starts its vision-language backbone from Florence-2 and learns the other components. Visual information receives positional features tied to a physical coordinate frame within each interaction window. With depth available, observations can be grounded directly; without it, the model learns to weight possible positions along calibrated camera rays. Camera calibration and robot-state information remain available in the paper's no-depth experiment. Calling this an ordinary, uncalibrated camera solution would overstate the result.

A trajectory tokenizer compresses a point's movement across time into one representation. The system then learns both motion and scene changes alongside robot commands. At execution time, however, its actions-only mode can skip generation of future action-flow and scene-flow outputs. The model can benefit from that training supervision without calculating every possible output for every command.

The authors measure 55.77 milliseconds per action chunk on an NVIDIA GeForce RTX 5090 D. Generating actions plus both flows takes 561.60 milliseconds with 3,000 scene queries. Those are different workloads on a specified GPU. Neither number establishes the end-to-end reaction time, power budget or safety behavior of a battery-powered humanoid.

Source: UNITAS v1, Tables 1–2. These are the authors' reported comparisons under their evaluation protocols, not independent Biped.News replications or a universal ranking of robot systems.

Scene prediction has its own scorecard. On held-out RoboTwin rollouts, the paper reports average moving-point displacement error of 0.427 centimeters, versus PointWorld's 0.656 centimeters. The often-cited reduction of up to 49% concerns static-region drift, not a 49% improvement in physical task success. Predicting geometry accurately and completing a manipulation task are related but separate tests.

Who's Building UNITAS

The author team includes researchers affiliated with DexForce, the Chinese University of Hong Kong, Shenzhen, Foshan University, Sun Yat-sen University and South China University of Technology. Ruixiang Wang and Yongyi Su are identified as equal contributors. This is a company-and-university research effort; the paper's results should be understood as the team's own evaluation.

The work sits inside a commercially funded robotics business. An October 15 Reuters report says DexForce raised 1 billion yuan in a recent Series B at a valuation above 10 billion yuan. That financing is company context, not independent validation of UNITAS or evidence that its research model is running at named customers.

The official repository labels its code Apache-2.0 and, as checked October 16, marks model checkpoints and inference/evaluation code as released. Training code and the training dataset remain unchecked in its release checklist. Download links and evaluation instructions make the claims more testable, but this is not yet a complete release of everything needed to reproduce training. No independent reproduction is established here.

What This Means for Humanoid Manipulation

For a humanoid team, the attraction is a shared representation that could connect human demonstrations, robot arms and different gripper geometries. It attacks a recurring integration problem: training data arrive with incompatible viewpoints, sampling rates and physical control formats. A useful common representation could reduce some adaptation work. That is an engineering implication of the design, not a measured saving in this paper.

The benchmark gap supplies a better buying signal than the largest percentage alone. UNITAS reports 99.8% on LIBERO, 89.1% on the perturbed LIBERO-Plus suite and 56.2% success on VLABench. The reported main-model setting uses mixed-data pretraining followed by downstream fine-tuning; LIBERO-Plus tests the LIBERO checkpoint without additional adaptation. These benchmarks pose different challenges, so their scores are not directly interchangeable. Together they show why near-perfect performance on one familiar test cannot stand in for broad competence.

The physical evaluation is narrower still: arranging books and placing a ball in a bag before closing its zipper, with 20 trials per method per task. A single trial changes a reported task rate by five percentage points. The study does not establish customer uptime, intervention frequency, behavior over entire shifts or handling of the many objects outside these tasks.

What's Coming Next

Three concrete developments would strengthen this result: an author clarification of the conflicting physical-task percentage; independent runs of the released checkpoints under the published benchmark conditions; and release of the remaining training code and data. Reproduction should preserve the distinction between pretrained-and-fine-tuned models and models trained directly on a target dataset.

For humanoid adoption, the next useful evidence would extend beyond tabletop completion rates: different hardware, longer task sequences, measured human interventions and documented recovery after failed grasps. A desktop-GPU latency measurement also needs an on-robot counterpart before it supports practical compute or battery claims. The authors give no deployment timetable that turns those tests into a promised commercial release date.

Frequently Asked Questions

A More Testable Route to Spatial Robot Learning

UNITAS makes a concrete contribution to the manipulation debate: actions and their consequences can be learned in an explicitly shared physical space, with published comparisons and tools that invite scrutiny. The limits are equally concrete: narrow physical testing, an inconsistent reported percentage and an incomplete training release.