TX-34 NOVA · What the campaign taught us
These are results, not papers: measured in simulation, reproducible with one command, and not submitted to anyone. What follows is what we found — including the part where we were wrong.
The single most useful thing this campaign found, and it was found by measuring a part nobody had measured.
29.5 cm
The foot we had
never measured from the model
22.5 cm
A real size 36
16.5 forefoot, 6.0 heel
+9.4 cm
Centre of pressure
ahead of the ankle
24
Configurations searched
for a forward twin
For three weeks the model walked backwards and every attempt to reverse it was aimed at the controller. The cause was geometric: the foot in the model was 30% longer than a real one, the centre of pressure sat 9.4 cm ahead of the ankle, and the residual moment drove the body backwards. The foot was rowing.
Correcting the geometry to an anatomical size 36 flipped the gait forwards. But the trade is exact and it goes both ways: the long forefoot that reverses the direction is the same geometry that sustains a long support. The backward mode covers 26 metres with a 12.3 cm stride — a human-scale stride, in the wrong direction. We searched twenty-four configurations for a forward equivalent. It does not exist.
The same 6 cm heel that comes with that foot is what limits squat depth, and the shortened forefoot costs 42% of the forward balance envelope and 27% of jump clearance. Every geometric decision on a biped is a trade, and this campaign measured both sides of four of them.
Fifty configurations across thirteen parameters. This is the question we are handing to a partner.
0
Actuators saturated
in any run that falls
8.1 cm
Centre of mass
outside a 4.5 cm foot
0.36 s
Gait cycle
in 53 of 53 configurations
100 g
Payload that
brings it down
The walking gait sits at 2 cm/s, and it is not a power problem. In every configuration that falls, no actuator reaches its limit. What fails is the weight transfer: when the next step begins, the centre of mass is still 8.1 cm outside a foot 4.5 cm wide.
The reason is one missing term. In support, the ankle law has damping and no position term — nothing is bringing the body over the foot. Adding one destabilises, because two loops then regulate the same quantity. To complete the transfer you need a longer support; to afford a longer support you need to have completed the transfer.
We also tried a predictive planner over three step parameters. It bought +71% speed and lost 47% of survival time — and the failure mode never entered a 1.5 s horizon. The fall of this platform is slower than the horizon that could avoid it, and that is a property of the machine, not of the planner.
Nobody has published this for a biped built on catalogue actuators.
38.2 N·m
Postural
89% of rated
29.8%
Of the cycle
contact transients
64%
Of the heat
from those transients
53 N·m
Touchdown target
to return to rating
The transient peak is governed by position-servo stiffness, not gait kinematics: raising the torque ceiling to ±200 N·m saturates that limit too, and the gait degenerates. The 80 N·m cap acts as a design limiter, not as a measurement of a requirement.
And the floor is honest: eliminating contact transients entirely would still leave the actuator at 89% of rating. The remaining lever is the lateral lean — a control strategy, not an actuator size.
A 5.5% mass increase does not degrade a platform uniformly. It destroys one regime and leaves another untouched.
Ballistic — jump height: −54%. Contact — landing pipeline: −100%. Quasi-static balance envelope: unchanged by mass. And at unchanged controller tuning, a 3.3% mass change cost 42% of the achieved support events.
Walking payload is a separate case and it is worth stating precisely: it is now zero, and not because of mass. The frontal hip actuator is at its torque cap unloaded, so a hundred grams is enough — and raising the cap does not help, which we know because we swept it from 60 to 140 N·m.
The ballistic regime is mass-sensitive; the support polygon is insensitive to mass. It is not insensitive to its own geometry — shortening the forefoot cost 42% of the forward envelope. That distinction tells a designer exactly where a change will hurt, and it is the honest answer to the first question anyone asks about a model with no CAD behind it.
Power loss on non-backdrivable gearboxes, three passive strategies compared.
Coasting leaves the head at 3.64 m/s at impact and takes 14.3 s to settle. Braking settles in about 1.5 s but puts the head at 5.3 m/s — because the joint stops turn the collapse into a rigid fall, like a plank rather than a body folding.
The earlier figure of 1.51 m/s published here is withdrawn: it does not reproduce, on any model version.
No passive strategy satisfies both criteria. It needs an active crouch or selective per-joint braking. This applies to any humanoid with non-backdrivable reduction, not only to this one.
This is probably the most transferable part of the work, and none of it was learned in advance. Each rule exists because something failed first.
Rule 01
Five times a headline figure changed for the same reason: the denominator contained time that did not belong to the phenomenon. Hip roll RMS was 43.1 N·m over the whole run and 53.7 in the steady-state window. Energy per step, being an integral over completed steps, did not move — which is how we knew the window was right.
Rule 02
Every module prints the mass of the model it actually loaded, with an assert. It exists because a whole batch once ran at 64.13 kg instead of 66.23, and the mass check — run separately from the tests — reported the right value of a different file. A check that runs apart from the test does not protect the test.
Rule 03
A seat reaction of 712 N on a body weighing 650 N. Impossible in quasi-static equilibrium. Cause: summing the magnitudes of contact forces instead of the vertical component, over a window placed in the descent. Every measured force now carries its closure check as a permanent assert.
Rule 04
The model was running with the hip 12° beyond the software limit. Once aligned, squat depth went from 0.48 to 0.60 m — the true value was worse than the published one — and sit-to-stand went from pass to fail before a different strategy recovered it.
Rule 05
Posture tables declared poses "verified against joint ranges". That is a kinematic check: it says the joint reaches the pose. Simulation found reachable poses that fall over.
Rule 06
An automated consistency checker declared the corpus clean while two withdrawn claims were still presented as achieved. Two causes: the forbidden-value list had been built from the memory of what was changed rather than from the list of dead claims, and exemptions had been widened until "clean" was reached. Now the tool must prove it can find — deliberately injected values, before any clean result counts.
Rule 07
On two hundred pages, two automated passes produced two new defects while correcting others: a retirement note made self-contradictory by a substitution applied inside it, and a withdrawn claim made more credible by replacing "certified" with "computed in simulation". Past a certain size, correct one at a time and then freeze.
Rule 08
A fixed joint between base and pelvis was being merged into the world by the exporter: 5.4 kg of pelvis evaporated from every URDF ever produced — 57.4 kg declared against 62.8 real — and no tool flagged it. Every derived artefact is now re-read and checked for mass and joint count, not just generated.
Nine figures withdrawn
The landing pipeline (0.36 s to upright rest) and the toe-brake multiplier (×3.9) did not survive a mass correction. The hallux contribution to jump height (+150%), the power-loss head velocity (1.51 m/s), the walking payload (2 kg), the static payload (6 kg), the forward balance envelope (0.60 m/s), the jump canon (210 ms, +3.3 cm) and the claim of cross-platform figures identical to the last digit all fell for different reasons — and each is declared with the measurement that replaced it. The jump was retracted twice: the second time because the ankle term named “push” was in fact dorsiflexion. Inverting the sign raised the centre of mass 34% — from 5.3 to 7.1 cm — while the feet clearance fell from 8.2 to 6.0. The old term flicked the feet up without lifting the body.
Almost none of them were arithmetic errors. They were correct numbers measuring the wrong quantity: a window that included the robot standing still, a contact counted as a step, a coefficient produced by reading the wrong actuator, a jump height measured from a height the robot leaves before it jumps. That failure mode is more dangerous than a bug, because it looks like a result — and finding it needs a second person, not a second run.
They are declared in the repository README, in the technical document, and here. A campaign that does not publish its retractions is not worth checking.
Formulated as questions, not requests. If you have data on any of these, it would be useful.
Every figure on this page is reproducible with one command from github.com/SynthMike34/nova-sim — identical on Windows and Linux, to the last digit. No result has been validated on physical hardware, and none of this has been peer reviewed. It is offered as it is: measured, declared, and open to being checked.