Many functional coatings, such as metallic lusters on tile or conductive traces on glazed ceramic, are made by firing a painted precursor in a kiln above 1000°C. That rules out anything already built into a floor, wall or facade. A recent laboratory result shows that the furnace can be replaced by a pulsed infrared laser. The laser decomposes a metal-organic ink (copper or nickel salts mixed with glass powder) into metal nanocrystals sealed in a melted glassy layer, in open air at room temperature. The catch is that the process only works inside a narrow window of laser power, scan speed and focus. Our proposal is to carry this laser process to installed surfaces on a legged mobile manipulator: a four-legged robot with an arm. The robot would treat the laser dose, not the arm pose, as the quantity its control system must hold steady while it walks.
The idea in two sentences is this. In this task the main disturbance flows from the legs to the tool: the walking body sways, which moves the laser's focus. So the leg controller should forecast its own sway and send that forecast to the tool, and the tool should cancel it by steering the beam and adjusting laser power so the delivered energy per area stays inside the process window. The tool also reports back which directions of sway it cannot absorb, so the legs learn to limit exactly those.
Core Claim
Existing controllers that split a legged manipulator into a leg module and an arm module pass information one way, from arm to legs. Combining Learning-based Locomotion Policy with Model-based Manipulation for Legged Mobile Manipulators feeds the forces and torques the arm will exert over the next 0.8 s (together called the predicted wrench) to a learned walking policy. Mobile-TeleVision feeds a learned summary of the arm's upcoming motion. Both treat the arm as the source of disturbance and the legs as the absorber.
Laser activation reverses this. The arm carries a light scan head and exerts almost no force on the body. What ruins the process is the gait-driven motion of the body itself, which changes the distance from lens to surface (the standoff) and the ground speed of the beam spot.
We claim that the right link between the two modules here is a short-horizon forecast of body motion from the legs to the tool, plus a "tolerance budget" from the tool to the legs. The budget is a small weight vector saying which directions of body motion the tool cannot compensate. We further claim that doing the compensation in the space of laser dose is what makes continuous processing while walking feasible. Dose here means approximately
$$F \approx \frac{P}{v\,w(\Delta z)},$$
where $P$ is laser power, $v$ is spot speed over the surface, and $w(\Delta z)$ is the effective spot width at focus error $\Delta z$. Holding $F$ within a calibrated band $[F_{\min}, F_{\max}]$ is much easier than mechanically holding the arm pose to that tolerance.
The process window in Processing and optimization of laser-activated Cu and Ni inks motivates this framing. At 2000 mm/s, 12 W leaves the precursor incompletely reduced, 24 W gives stable, vivid results, and 36 W or more causes plasma, shock waves and ablation. Power can be changed at kilohertz rates, far faster than any leg or arm can move. The paper was run with a fixed galvanometer head (fast steering mirrors) over static tiles, so the tolerance of this window to focus error is unknown. Mapping it is part of the work.
Research Program
The first stage is process characterization on a bench. We would reproduce the Cu and Ni inks and the 1064 nm pulsed fiber laser conditions reported by Lahlahi-Attalhaoui et al.. We would then extend their power-by-speed sweep with a third axis, deliberate defocus. This gives an empirical map from $(P, v, \Delta z)$ to outcome. Outcomes are classified as unreduced, good, or ablated, using the reported optical signatures: the copper absorption band near 450 nm and the nickel band near 350 nm, which come from surface plasmon resonance, the collective electron oscillation in metal nanocrystals. X-ray diffraction checks the metallic crystal phase on sampled points.
We would fit $w(\Delta z)$ and the dose band from this map. We do not assume the simple formula above holds; if outcomes do not collapse onto a single dose variable, the tool controller uses the full three-dimensional map as a lookup instead.
The second stage builds the tool-side controller on the arm. It includes a galvanometer scan head and a standoff sensor, a fast mirror loop that corrects lateral spot error, and a power loop that holds dose within the band. The power loop combines the sensed standoff with the legs' forecast of body motion, so it can act before the error is measured.
The third stage trains the leg policy in Isaac Gym (Isaac Gym), following the reinforcement-learning recipe of the arm-to-leg forecasting work. The policy gets two new ingredients. It outputs a forecast of body pose over a short horizon, trained with a supervised loss against the body pose it actually produces. It also receives the tool's tolerance budget as an input, and is penalized for body velocity along those weighted directions.
Hardware is an ANYmal-with-arm platform of the kind used in that forecasting work, or a Spot with arm as used by ReLIC. In both cases the laser source is carried on the body or tethered by fiber, under an interlocked enclosure.
The application target is patterning metallic decoration or traces onto pre-inked installed floor and wall tile, in areas larger than any fixed scanner field. Kiln firing is impossible there, which is the scalability argument.
Baselines
The naive reference is blind decoupling: a standard walking policy with no knowledge of the arm or tool, the arm on inverse kinematics (solving for joint angles that reach a target pose), and laser power fixed at the bench optimum of 24 W. This is the configuration the forecasting work reports as "naive"; its forecasting controller improved lateral force tolerance by 108% over it. It shows how badly unmodified walking sway degrades the process. It runs on the same robot, ink batch and path set.
The strongest existing method is arm-to-leg wrench forecasting as in Combining Learning-based Locomotion Policy with Model-based Manipulation. That paper reports a 57% reduction in end-effector position deviation relative to a walking-only base, and resistance to 150 N lateral force without stepping. It is the best published way to keep an arm tip steady on a walking base. We would retrain it in the same simulator, with the same reward-tuning budget and number of environment steps as our policy, and pair it with the same galvanometer and fixed-power tool. To beat it, our system must raise the fraction of scanned path length classified as good by a margin larger than run-to-run variance, at equal walking speed.
A monolithic baseline trains a single policy for all joints, following Deep Whole-Body Control. That paper reports an 87% survival rate under perturbation and a 0.82 m³ workspace versus 0.58 m³ for a decoupled controller. Here the single policy observes standoff and outputs laser power as an extra action, trained with equal compute.
Two ablations isolate our key ingredients. The first keeps dose compensation but removes the body-motion forecast, so power reacts only to sensed standoff; this isolates whether prediction beats sensor latency. The second keeps the forecast and tolerance budget but fixes power, which isolates dose-space compensation.
No reported result was found on Lacuna for laser processing from any mobile or legged platform, so the process-quality numbers for all baselines will be new measurements.
Why not just…?
Why not just stop the robot, stand still, and scan a patch? This is the most serious alternative, and it may well win on quality for small areas: a standing robot sways little, and at 2000 mm/s a patch scans in seconds. Its costs are the seams between patches and the time spent stopping and restarting. We will measure both directly with a stop-and-scan mode. We will report good-path fraction, seam defects and area per hour against continuous walking. If stop-and-scan matches our throughput within noise, we will say that continuous processing is not needed and that the contribution reduces to the process-tolerance map.
Why not just add autofocus to the scan head and ignore the legs? Purely reactive tool-side correction is exactly our first ablation. The question is whether standoff sensing latency and the limited range of focus correction can keep up with body sway during walking. Mobile-TeleVision found that reacting without anticipating body motion leaves large errors, which suggests prediction matters, but that was for arm pose, not focus. If the ablation matches the full system, the forecast channel is unnecessary, and we will report that.
Why not just train one end-to-end policy? Unified policies can find whole-body coordination on their own, as Deep Whole-Body Control shows with its larger workspace. For forceful humanoid tasks, however, FALCON reports that splitting control into two agents doubled upper-body tracking accuracy over both monolithic and IK baselines. And a laser power action inside a policy trained in simulation would depend on a simulated ink response, which does not exist. The monolithic baseline above tests this directly.
Why not a wheeled base or gantry? For flat, open factory floors, a wheeled base or gantry is better, and we do not claim otherwise. Precision gantry control of a deposition process is already demonstrated, for example in Force Controlled Printing. The legged case is justified only for stairs, thresholds, cluttered renovation sites and walls reached from uneven ground. Our field trials are chosen to include those conditions.
Proposed Experiments
The first decisive check is on the bench. Does laser outcome collapse onto a single dose variable across defocus, and how wide is the focus tolerance at 24 W and 2000 mm/s? A window of only a few hundred micrometres would make power compensation the core mechanism. A wide window would make the whole problem easy and shrink the contribution.
The second check is controlled walking over pre-inked tile laid on flat floor, a 10° ramp and a step. On each surface, all baselines and ablations scan identical line and fill patterns at matched walking speed. We measure the distribution of focus error, the deviation of delivered dose from the band, the good-path fraction from reflectance mapping, ablation and unreduced defect counts, and adhesion and durability checks on the vitrified layer.
The third check is scaling. We compare area per hour and seam quality between continuous walking and stop-and-scan over tiled areas several times larger than the scanner field.
The proposal succeeds if the full system beats the retrained wrench-forecasting baseline on good-path fraction, and beats stop-and-scan on throughput at comparable quality. Each ablation must lose measurably; otherwise its ingredient is dropped from the claim.
References
- Combining Learning-Based Locomotion Policy With Model-Based Manipulation for Legged Mobile Manipulators
- Mobile-TeleVision: Predictive Motion Priors for Humanoid Whole-Body Control
- Processing and optimization of laser-activated Cu and Ni inks for ceramic tile inkjet printing
- Isaac Gym: High Performance GPU Based Physics Simulation For Robot Learning
- Versatile Loco-Manipulation through Flexible Interlimb Coordination
- Deep Whole-Body Control: Learning a Unified Policy for Manipulation and Locomotion
- FALCON: Learning Force-Adaptive Humanoid Loco-Manipulation
- Force controlled printing for material extrusion additive manufacturing