Deployment objectives often specify desired outcomes without providing revised trajectories, while parts of the original reference must still be followed. We present OneSteer, a one-step diffusion whole-body controller that combines general motion tracking with selective test-time steering. To accommodate new objectives, differentiable task costs guide optimization of the generator's input noise with frozen weights, while body-group reference masks relax selected tracking requirements. Predicted state-action plans provide a shared interface for task optimization and feedback-conditioned execution. For nominal tracking, the model trained from scratch achieves the lowest local tracking error among the evaluated baselines on successful held-out AMASS and BONES clips. For steering, one frozen distilled controller adapts across diverse deployment objectives in simulation, raising the clean passage rate under a 1.25 m bar from 2% to 90%. On a physical Unitree G1, OneSteer completes all 27 nominal tracking trials across nine diverse whole-body motion sequences without falling. With a steering cost derived from a signed distance field built using only onboard LiDAR, it also passes under a 1.1 m low bar in four of five attempts and steps over a 12 cm hurdle without falling in all three attempts. Beyond control, prediction variation in the model trained from scratch ranks likely tracking failures without physics rollouts. This scoring method is more than 10x faster than simulation-based scoring with five rollouts per clip and guides targeted fine-tuning. Project page: https://onesteer.github.io
Alex: A robot can copy a walk perfectly and still walk straight into a low bar. OneSteer lets a humanoid change that movement for a new task without retraining, while preserving selected parts of the original.
Sam: So the goal is not just faithful imitation. It is knowing what to change when imitation would cause trouble.
Alex: That is the problem Yifei Huang, Yazhao Yang, and colleagues tackle in OneSteer, a methods preprint. They test a controller in simulation and on a physical humanoid.
Sam: For researchers building robot behaviours, that could mean fewer separate policies for every obstacle or load. But adapting a movement is useful only if the robot can actually perform the adaptation.
Alex: That distinction drives the study. We will follow how they change the movement, why ordinary trajectory editing falls short, and where the evidence stops.
Sam: Start with the existing approach. If a robot already follows recorded movements, why not edit the recording to avoid the bar?
Alex: That is an established option: edit the reference motion, meaning the movement the robot is asked to follow. A planner adjusts the poses, and a separate tracker tries to execute them. The authors argue that the gap between those jobs can defeat the edit.
Sam: A crouch can look correct in the planned movement, while the robot barely lowers itself. Do they test that gap directly?
Alex: They do. Their edited references clear the bar, but the trackers often fail to reproduce the requested lowering. OneSteer instead changes a plan inside the controller that will execute it. The central claim is not merely better planning; it is better translation from an adjusted plan into movement.
Sam: Generative controllers already produce movements. What does this one add beyond that family of methods?
Alex: The authors distinguish those controllers by what they follow. Prior integrated systems use inputs such as sparse goals, keyframes, or text, rather than a sliding full-body reference window. OneSteer combines general reference tracking with test-time steering: changing behaviour during execution without changing the network’s weights.
Sam: Walk us through the mechanism with the low bar. What information enters the controller?
Alex: It receives upcoming reference poses and proprioception, measurements of its own body state. A generator turns those inputs, plus noise, into a short plan of future states and actions. The authors call it one-step diffusion because generating that plan needs no iterative chain of denoising operations.
Sam: The input includes noise, but the objective is controlled motion. How does that become a useful adjustment handle?
Alex: They define a differentiable cost: a score whose gradient tells them how to change the input. For the bar, the score penalizes predicted body positions that enter the obstacle. They follow that gradient back through the generator and adjust its input, leaving the trained weights fixed.
Sam: So they are not directly dragging the robot’s head downward. They change the generator’s input so it produces a different whole-body plan.
Alex: Yes. There is a practical wrinkle: one-step describes each generation, not the entire steering procedure. Steering uses a draft plan, mixes it with fresh noise, then refines that input before generating again. The draft-containing input responds better to guidance than pure noise does.
Sam: But the original reference still says to walk upright. What stops the controller from undoing the crouch?
Alex: They divide the reference into root, legs, and upper-body groups. Masks hide selected groups, releasing those tracking requirements. Crucially, a conflicting target must be hidden from the action-producing head, not just the generator. Otherwise, the head pulls the movement back toward the reference.
Sam: That makes the mask a handover: stop insisting on this part of the original movement, and let the task guide it.
Alex: With a qualification: the visible groups remain tracking targets, not guarantees. The head uses near-term planned states, current feedback, and the remaining reference to produce the actual command. The robot applies that command, measures its new state, and replans.
Sam: Let us separate the evidence. Does making the tracker generative reduce its ability to follow ordinary movements?
Alex: The authors train separate versions by different routes. One learns from an existing reinforcement-learning teacher, MimicLite; another learns from scratch using a learned model of robot dynamics. On successful held-out clips, the scratch-trained version has the lowest local tracking error among the evaluated baselines on both motion benchmarks.
Sam: That is a tracking result for the scratch-trained model. Which version supports the main steering comparisons?
Alex: The distilled version, learned from MimicLite. That allows comparison against editing the same teacher’s references. Keeping those versions separate matters: the best tracking result and the headline steering result do not come from the same trained model.
Sam: What is the clearest steering result, and how broad was the test?
Alex: They tested seventy-nine walking motions from unseen actors. Under the main low bar, clean passage rose from two percent without steering to ninety percent with steering. Clean means clearing the bar without collision, not going around it.
Sam: Beating an unchanged walk is useful, but the edited-reference controller is the stronger comparison. How did that perform?
Alex: The best editor reached only eleven percent clean passage under that bar. Even letting the teacher track OneSteer’s own adjusted plan fell short of OneSteer’s full controller. That supports the authors’ argument that the execution head contributes, not just the plan.
Sam: Does that advantage hold equally across all the tasks?
Alex: No. The strongest comparative evidence is under the lower bars. The hurdle advantage over closed-loop editing is less decisive at the motion-window level. Reaching can improve with more optimization, but the larger computation budgets exceed the real-time settings.
Sam: What happens when the simulated robot becomes a physical one?
Alex: On a Unitree G1, steering passed under a low bar in four of five attempts. In the failed attempt, it crouched too far and its waist touched the ground. Hurdle trials avoided falls, but most crossings still involved foot contact.
Sam: Those attempts establish feasibility, not dependable deployment. Was the obstacle sensing and computation fully onboard?
Alex: The obstacle information came from onboard laser scanning. It produced a signed distance field, a map measuring distance inside or outside obstacles. But the map was built before each run and kept fixed; steering ran on a tethered computer. Onboard timing tests were separate from those hardware demonstrations.
Sam: There is also a deeper limitation: improving the predicted movement does not guarantee improving the executed one.
Alex: The authors document that gap. Guided plans can change joint positions without matching velocity changes, and the head executes some adjustments only partly. A learned dynamics model helps repair some plans, but adds computation. So this is not a guarantee of collision-free or dynamically consistent motion.
Sam: Beyond obstacles, is the frozen controller genuinely reusable?
Alex: In simulation, the same weights adapt to weakened motors, tray balancing, reaching, and velocity commands through different costs. The paper also uses disagreement between generated plans to flag likely tracking failures without simulation rollouts. That is an empirical ranking signal, not a calibrated failure probability; agreement can still hide a biased plan.
Sam: Who should read the full document, and where would you send them first?
Alex: Researchers in humanoid control and generative control should read the comparison with kinematic reference editing first. Then read Test-Time Steering for the masks and generation procedure. Follow it with the appendix discussion, Executing guided plans, which shows where the attractive plan still outruns the robot.
Sam: And for someone who only needs the central lesson?
Alex: A movement is not successfully adapted until the controller can execute the adaptation. OneSteer makes progress by changing the plan inside the system that acts on it, while keeping the remaining reference in view.
Sam: That is a useful distinction to carry into the next robot demonstration.