RoboHarness: A Simple Harness Outperforms VLA and World Action Models

RoboHarness. A Simple Harness Outperforms VLA and World Action Models. By Bince Qu and two others, from Zhejiang University. Published on the sixteenth of October. Its code repository is worked out: about seven hundred source files, with a readme and a licence.

Large language models have opened a path toward generalist embodied AI, and with it a practical question: how should such a model act in the physical world? One straightforward answer is to train the model to generate actions itself. Collect robot-action data and video demonstrations, refine the language model's architecture with an action head, and train a vision-language-action model, a VLA, at scale. That route asks for a great deal of data, architecture work and training before the model can do anything.

The authors argue for a different route. Coding agents, they say, already have the potential to realize generalist embodied manipulation. What they lack is not more training but a simple and effective robot harness: an interface between the agent and the robot, rather than a change to the model.

RoboHarness is that harness. It gives language model agents a visual-geometric control panel, so that they can directly understand embodied tasks and invoke them. The agent stays the agent it was. What changes is the surface it acts through, one built to present the physical task in visual and geometric terms that the agent can read and call into.

The test is the BEHAVIOR Challenge 2025 tasks. Running through RoboHarness, Qwen3.8-Flash-Next outperforms state-of-the-art VLA models on those tasks. The implementation is open source.

The claim here is about where effort goes. Instead of teaching a language model to emit actions through large-scale robot training, the paper hands an existing coding agent the right control surface, and reports that this alone is enough to beat models trained for action on a shared set of embodied tasks.

RoboHarness. A Simple Harness Outperforms VLA and World Action Models. By Bince Qu and two others, from Zhejiang University. Published on the sixteenth of October. This is the long version: the whole paper rather than the abstract.

A large language model knows a great deal about the world, but that knowledge lives in text. A robot needs motion. The usual bridge is to train the model to emit actions directly: a vision-language-action model, or VLA, which either writes actions as tokens or hangs an action head off the language model. That can express any motion, but it needs robot-action data that is hard to collect at scale, plus extra post-training. World action models, which learn how the world changes under an action, need training too. The older alternative gives the model callable skills such as pick and place to compose into programs. The trouble there is coverage: every new object or task may need another function. The authors, from Zhejiang University, read this as a problem with the interface, not with the model.

Their starting point is a lesson borrowed from coding agents: a minimal set of simple, orthogonal primitive tools beats a pile of predefined features and heavy scaffolding. So RoboHarness trains nothing. It plugs into coding-agent harnesses such as Claude Code and hands the model what the authors call a visual-geometric control panel. Two observations make that plausible. First, the model only has to mark points on an image, which grounding training makes it reasonably good at. Second, a manipulation task can be viewed as two stages: the gripper grasps a key, say, and then the held key is driven to meet the insertion constraint with the lock. If both stages are constraints between points, a small toolset covers them.

Three pieces do the work. The first is one camera frame shared by every robot, with forward, left and up as the axes. The second is a tracking tool. The model marks points in the head camera view; the harness follows each with optical flow, samples its depth, and back-projects pixel and depth into a position in the robot frame, which the model keeps receiving as the scene changes. The third is the action toolset. The model names points and the relations they must satisfy, separating on-hand points, on the gripper or the held object, from off-hand points fixed in the scene. Common relations come as quick requests: touch, where two points coincide, linear, where points are collinear, parallel planes, and a few more. The harness then samples candidate poses, keeps those that satisfy the constraints, discards any whose inverse kinematics are infeasible, plans a trajectory and executes it. When constraints fail, for instance because an object slips out of the gripper, the model can fall back on adjustment tools that nudge an end-effector, the chassis, the body height or pitch directly.

A proof is why the toolset is called complete, not merely small. The gripper carries three tracked points: the two fingertips and a centre point. The first proposition shows that any contact pose maps one to one onto three surface points: each fingertip overlaps a surface point, and a third surface point is collinear with the midpoint between them and the gripper centre, which fixes the remaining rotation. The second shows that, once an object is held rigidly, any relative placement of it against another object is pinned down uniquely by two non-parallel line coincidences. The first line leaves only a spin about it and a slide along it. The second line fixes both.

The headline test is the first ten tasks of the BEHAVIOR Challenge two thousand twenty-five, long-horizon household jobs on the R1 Pro robot. RoboHarness runs the open model Qwen3.8-Flash-Next. Averaged over the ten tasks, it scores forty-four point five percent, against thirty-four point eight for Galaxea G0.5 and thirty-one point six for RLC, the challenge winner. The pattern is uneven. It loses clearly on turning on the radio, where Comet scores a perfect one against its zero point four, and on Christmas decorations, where pi zero point five more than doubles it. Can meat defeats every system.

The more direct test of the thesis is the generalization study: pick up each of one hundred rigid objects, a success being a lift of at least five centimetres within two thousand steps. The RoboHarness runs, with zero hours of robot data, score sixty-eight to eighty-six, above the learned policies. Among those policies, robot data does not predict success. Faster-WAM reaches sixty-one from about three point eight six hours of demonstrations, while several models trained on thousands of hours score lower. Model size does not predict it either: the smaller world action models outscore larger VLAs such as OpenVLA-OFT. The authors conclude that multimodal capability matters more than parameter count.

There is a short limitations section, candid about the main caveat. The model plans well but often fails to ground plans in the robot's body, knocking objects over, falling or getting stuck. So on BEHAVIOR the authors hand-wrote a prompt for each task, and the appendix says it includes a procedure for accomplishing the task. Part of the planning credit belongs to a human. The evidence is also thin by the paper's own account: five instances per task from one seed, because of the evaluation budget. There are no error bars, and scores swing from zero to one within a single task. The completeness proofs assume a rigid grip, and they show the constraints can express a pose, not that the model will choose the right points.

The idea to carry away is where the boundary between reasoning and control is drawn. Instead of teaching a language model motor commands, RoboHarness lets it point at things and state geometric relations, leaving tracking, kinematics and planning to the harness. A proof says that vocabulary can describe any grasp and any placement of a held object. The experiments suggest an untrained model using it can beat trained policies on average, though for now a human still outlines each long task.