Compensating a humanoid nailer's backlash works only if the compensator knows the balance controller

Predicting each held joint's side removes a G1's backlash offset only with the robot's own balance controller in the model. A wrong one leaves 1.2 to 1.9 mm.

The buildbot-sheathing programme asks whether a robot can hang a 4x8 sheet of wall sheathing and land all 62 nails inside their windows on the studs. Two earlier steps found that gear backlash in a humanoid's joints leaves a fixed offset at the nailer's nose. The offset changes from pose to pose along the wall, so one tool calibration can't remove it. On a Unitree G1 leaning into the wall to take the press, one calibration still leaves 1.1 to 2.6 mm. Both steps named the same remedy: use the static torques to predict which side of its gear gap each loaded joint rests on, at every pose, and subtract that bias. This step tests the remedy.

The setup. The poses are the previous step's: 15 for the G1 and 21 for the H1, hand heights 0.4 to 1.2 m (H1 to 1.6 m), each straight ahead and shifted 0.15 m either way along the wall. The robot carries a 3.4 kg nailer (the Metabo HPT listing) and presses 25 to 50 N. The scoring is the programme's best single-2x pipeline with 12-bit joint sensors and 5 arcmin of backlash per joint. The robot balances with one of two controllers. One minimises leg torque ('torque'). The other also keeps the light leg joints (hip rolls, hip yaws, ankle rolls) loaded one way ('hold-min'). The compensator predicts torques with its own model, which can be wrong in two ways: it can assume a different ground-reaction split, or it can miss mass at the tool. Three hypotheses were registered before the run.

With the right model, nothing is left. When the compensator uses the robot's own controller, the leftover offset is exactly zero at every pose. Each pose then scores what it would with its own calibration. Run with nothing predicted, the same code reproduces the previous step's 1.401 mm.

With the wrong controller model, compensation barely helps (H121 held). Take a G1 on 'hold-min'. A compensator that assumes the textbook least-norm split of the ground reaction leaves 1.156 mm. It puts 17 joint-pose cases on the wrong side of their gaps and misses 47. Assuming the force-minimising or torque-minimising split leaves 1.450 and 1.237 mm. Doing nothing leaves 1.401 mm. With the G1 on 'torque', the wrong models leave 1.24 to 1.92 mm, against 2.63 mm for doing nothing. The waist and arms carry the same torque under every split. The leg joints do not, and on these poses their gaps move the nose most.

A full nail magazine hardly matters, if the controller holds its joints (H122 refuted on its strict count). The magazine holds 64 nails. From steel density and nail size, that is 0.295 kg for 8d commons (2-1/2 x 0.131 in) and 0.410 kg for 16d (3-1/2 x 0.131 in). Leave 0.4 kg out of a correctly modelled compensator on 'hold-min' and two joint-pose cases flip, out of hundreds. That refutes the registered "no misprediction" claim, but the leftover is only 0.017 mm. A full kilogram, an allowance for the hose that isn't sourced, leaves 0.057 mm. Under the 'torque' controller the same error matters far more: +0.2 kg leaves 0.412 mm and +1.0 kg leaves 0.836 mm. That controller leaves light joints wherever the optimum puts them, often close to zero torque, so small loads flip them. 'hold-min' keeps them at least 0.2 N m from zero.

Skipping low-torque joints makes it worse (H123 refuted). The obvious hedge is to compensate only joints whose predicted torque clears a guard band, say 1 N m, and leave the rest alone. For the least-norm compensator that leaves 1.374 mm: wrong sides fall from 17 to 11, but missed joints rise from 47 to 79. A joint left out still sits at one side of its gap, so it costs half a gap, the same as guessing its side wrong. On the matched compensator a band turns zero into 0.96 to 1.35 mm.

What a builder would do with it. Pair a controller that keeps the light leg joints loaded one way with backlash compensation computed from that same controller: the same contact QP, the same held set, at every pose. A generic statics model of the robot is not enough. A tool that gets heavier as the magazine fills is then not a problem.

What it doesn't settle. This is statics only. The "real" robot here is still the MuJoCo model: its gains, joint friction and any asymmetry in the gear gaps are not measured, and Unitree publishes no backlash figure. The poses don't re-lean for the extra mass. One H1 pose has no 'hold-min' solution and was skipped. Nothing here was tested on hardware.

Code, results and figure: research/buildbot-sheathing, experiments/m33_compensate/. Project page: https://marisstella.app/research/buildbot-sheathing/

All writing · More in Research findings · More on buildbot-sheathing · All tools and research