Kinewright Humans: How a runner stops

Kinewright is a house of motion engines. Humans is a headless TypeScript engine that turns a body, a motion skill, a stream of commands, time and a seed into a deterministic stream of 3D poses and contacts. It outputs motion, not pixels. The picture in the videos is a plain Canvas 2D reference renderer; the engine is what moves it.

The level today is kinematic. There is no physics engine underneath: no forces, no ground reaction, no balance. Every number in this article is an engineering threshold that the motion was built to, not a measurement of anything.

Before this stage the engine could walk, run, go from a walk into a run and from a run into a walk, and both of those transitions had been accepted by us, the two people who look at every result. This stage asked for the rest of ordinary locomotion: standing, starting, stopping from a walk, stopping from a run, and commands that interrupt each other. The hard part turned out to be the last stop.

The video

18 seconds, real time, rendered at 60 fps on a fixed-step clock, nothing hand-animated; the GIF here plays it back at 15 fps. It was rendered from the exact bytes we accepted the stage on.

One thing to watch at about 14.9–15.3 s, on the last two braking steps of the run exit: the foot is put down hard. It reaches the ground while still rising and stops dead, where a walking or running foot comes down onto it. We accepted it as it is at this stage and will rework it later, in the step planner that places each foot from the body's speed plan and brings it down onto the ground. The rest of this article explains what you are seeing, in order, and ends with that foot.

Three short clips accompany the article, at 30 fps: a steady walk, a steady run, and stand → start → two steps → stop. They are the same engine on the same scene; the walk and the run are the previously accepted motions the new ones had to join.

Standing

A stand is a hold, not a pose with sway. There is no breathing and no micro-movement in it, on purpose: a living idle is a separate piece of work and it was not faked with noise. The stand is entered when a stop completes and left when the stop is released. It was checked to hold the same pose and the same two foot contacts for a five-second dwell at real time, six hundred frames, with no event and no drift.

Starting

A start is a weight transfer and then a step. The accepted start the engine already had took 0.8 s to prepare the first step and used three accelerating steps. We wanted a start that begins at once, so the product start prepares for 0.35 s and takes two accelerating steps, and the walk it settles into is the accepted walk unchanged. Movement begins 0.3989 s after the command against a bound of 0.4 s. That is a margin of one millisecond and it is stated as one.

The first version was refused before any formal look: it unlocked the legs and sat down before the first step, where a person simply begins with a step. The crouch and the return to neutral were removed. What was left after that was a 2.26 mm step down at the instant movement starts, the descent to walking height that the crouch used to carry. It is now blended across the first step, so the pelvis leaves standing height continuously and is at walking height when the foot lands.

That costs one thing, said here rather than left to be found: the planted foot floats 0.21 mm above the ground for about 0.13 s while the pelvis is still high. It is twice the tolerance the stage holds its seams to, nothing penetrates the ground, and we accepted it rather than hide it.

Later, a frame-by-frame check added for a different reason found two more defects in the start that nobody had seen: the root, the head and both hands stepped 20.6 mm sideways between two frames, once at each foot's liftoff. The cause was a membership test. A standing foot held the pelvis back by its reach as long as it was "on the ground", so when the foot left the ground the restraint left with it and the answer stepped. A foot now holds the pelvis back in proportion to the weight it still carries, and the jump is gone.

Two things a start does not do yet. A stop pressed during a start still costs a full stopping motion, seven foot contacts, where we wanted at most one more. And a tap of the walk command does not give exactly one step: it also costs seven contacts and reaches the hold 3.29 s after the release. Both are transferred to the next substrate, a runtime step planner, as not met. The reason is structural: the one route that could deliver a single step, a weight transfer, one step and a brake from that step's low speed, is entered at a foot contact that is still accelerating at 0.6 m/s, and this compiler lays out exactly one route with the stop always last.

Stopping from a walk

A stop is a motion request, not a pause. From a walk the engine enters the accepted braking motion at the nearest foot contact of either foot, half a walk cycle at worst, and brakes over two steps into standing. The braking is the accepted chain's own: 0.48 s down to 0.9 m/s, then 0.58 s to rest. What changed is only where it may be entered: at the contact the deceleration begins on, rather than a stride and a half before it.

The first version was refused before any formal look, as a jerk. It put the whole deceleration into one short step. The second version, entering the accepted deceleration where it begins, was the one we looked at.

We refused its ending. After the feet came together, both knees straightened and pressed down at once as one last movement, where naturally one leg finishes completely and then the other, with no such movement at all. The closing was rebuilt around that sentence: the supporting leg is straight at the last contact; the closing leg lands bent at 13.7° and straightens to 0.01° as it takes the weight; the pelvis rises inside the last step; and nothing moves once the feet are together. What is left after the feet meet went from 0.170 s to 0.065 s, and the torso arrives upright with the feet instead of unwinding for 0.18 s after them.

To make the supporting leg reach full extension before the contact rather than on it, the pelvis rolls toward the supporting leg across the last step, by that leg's own reach deficit, under a 3° ceiling. Two earlier attempts had returned the straight leg as geometrically impossible; the argument held for a level pelvis only.

On the next look we refused the upper body: the torso shook side to side and the shoulders swung too far, and from a walk into a stop it leaned to the right. The 3° of pelvis roll was arriving whole at the shoulders, 3.04° of shoulder line and 34 mm of head travel, where the accepted closing had 0.11° and 2.5 mm. The walk's own pelvic obliquity lives in the pelvis joint and the spine already counters about two thirds of it; the closing roll had been applied underneath everything. The spine now counter-rolls the closing's roll, split between lumbar and thoracic by a ratio read off the rig's own spine offsets, so the head stays over the pelvis as well as the shoulder line staying level: 0.057° of shoulder line against a bar of 0.62°, 1.28 mm of head against a bar of 12 mm. It can move no foot, because both hips hang from the pelvis and nothing above the lumbar can reach one.

One request from that look was declined with its price on the table. We asked for the pelvis to level after the contact across the closing leg's weight acceptance, about 0.13 s, rather than the 0.05 s it takes. It cannot, because the roll is the supporting leg's reach deficit and the deficit is what is left of the pelvis's rise; they are one motion. Forcing the window open lifts the supporting foot 1.8 mm off the ground for the whole of it. I read that and kept 0.05 s.

Stopping from a run

This is where the stage spent most of its time. We refused one motion six formal looks in a row, each time at a finer grain, and each refusal is in the video's history.

1. It walked first. The first version reused the walk's stop, so a running character went into a walk and walked about three and a half steps before stopping. We refused it: stop at once, and from a run go into walking and stop, not walk for a while.

2. A direct route. The engine now leaves the run at the run's own boundary into a dedicated run-to-stand motion: it decelerates on braking steps to the point where the gait becomes a walk, then plays the same two-step stop, with no steady walking in between. The worst case from the stop command to the hold is 4.46 s against a bound of 5.5 s.

3. The braking steps floated. We both refused the run's deceleration as it stood: it glitched, it was sluggish, a moon walk, a skipping gait. In numbers, the first braking step was 29 % longer in time and 15 % longer in distance than a run step and its apex 18.6 mm higher, and the second flew 49.1 mm above the run. The exit was recompiled so that braking steps keep the run's own cadence while the step length and the flight shorten with the speed; no braking step is longer or higher than a run step. One requirement is still missed, by 22 micrometres: the first braking step's apex sits that much above the run's, where it used to sit 22.3 mm above.

4. The leg "chewed". Our next look: at the end, the leg is chewed over in the last movements. What we were watching was the step interval jumping from the run's 0.329 s to the walk's 0.420 s in a single step, a 27.8 % jump, while the last braking step still carried a 62 ms flight. I replaced the requirement in my own words: the run's cadence holds while a step still ends at 3 m/s or above, and from there the interval grows step by step to the walk's own at the commit. The intervals are now 0.329, 0.357, 0.387 and 0.420 s, every neighbouring pair within 8.5 %, no flight on the last two braking steps, every apex under the run's. The finding behind it: the interval could not travel alone. The run's stance fraction bottoms out at about 0.405, so widening the interval without widening the stance makes the flight longer, not shorter, and the stance now tapers with it. The taper spends the same ground at longer intervals, so the exit takes 177 ms more and the worst case moved from 4.28 s to 4.46 s, still inside the bound.

5. The leg kept the run's shape. Next look: something is still wrong with the leg, maybe the foot. The timing had become a walk's, but through every braking step the limb kept the run's form: the foot pushing off its toes, the knee thrown past ninety degrees, the toe carried high, the planted foot travelling most of a walk step over its own stance. All four switched to the walk's form in one step at the commit. All four had the same single owner inside the compiler, an amount that says "this is a run" and was exactly 1 until the commit. That amount now crosses to the walk's on the same fraction the interval and the stance already travel on, and the pelvis descends with it, apex by apex, where it used to stay at running height and drop at the commit. One bar in this set I closed unmet, by decision: the planted foot still travels 25 mm over the last braking stance against my bar of 15 mm, because taking the rest would have cost the pelvis dipping under the walk's floor, which we had refused earlier.

6. "Like an old mechanism jammed." Next look: the last steps look very mechanical, as if an old mechanism seized. Everything else we approved. A whole-window frame-by-frame scan of every clip, added for this, found three jumps the existing checks had walked past because they only checked the boundaries between clips:

• the body crossed sideways at 5.69 m/s inside a 21 ms flight. The pelvis had been held exactly over its planted foot for the whole stance, so its crossing to the next foot's line had nowhere to happen but the flight. On the run's 106 ms flights that is 0.1 mm of sway; on the tapered exit's 21 ms flight it had grown to 77 mm;

• the whole body moved 48 mm in one tick, 77 ms into the first stopping step, taking the head, both hands and both knees with it. It had been in five looks and every check had passed it;

• the two 20.6 mm jumps in the start, described above.

Two of the three were the same shape: a set that decides an answer lost a member at a hard boundary, and the answer stepped. The crossing now opens a little earlier, 36 ms instead of 21, the widest the existing continuity ceilings pay for; the pelvis hands over from the foot-based support to its guide across the double support that is taking the weight; and the frame-by-frame check over the interior of every clip is permanent. Two things were priced and returned rather than taken: a braking curve whose peak equals its mean, because it costs 2.8 micrometres of a continuity ceiling I had set, and an even deceleration, because the ground the route has fixes the curve's exponent at about 2.23 and an even one would need 0.76 m more ground than the route has.

7. A jerk forward on the penultimate step. Next look: almost there, but the second-to-last step jerks forward with the leg before the foot comes down. Two things were found. From the third braking step the feet were landing 88–89 mm either side of the route, the standing track, where the run lands at 49.5 mm and the walk at 46.4 mm; the run's narrow track is released by an amount read off the speed, which had expired mid-braking, and the walk's narrowing is a correction a running step is excluded from, so nothing held the ground between the two. The pelvis had to travel to meet the feet: it swung ±71–75 mm and crossed between them at 3.46 m/s against 0.59 m/s in the run. The track now tapers from one gait's to the other's on the same fraction as everything else, and the pelvis crosses the route at 2.47 mm per frame, the accepted run's own figure, where it crossed at 14.3.

The second thing was not answered and the record says why in numbers. The pelvis does not follow its own speed plan through the exit. At the four braking contacts it sits 103, 149 mm behind and 59, 191 mm ahead of where the plan puts it, because the foot is placed from the plan by the step's own length while the leg's geometry is on its way from the run's to the walk's. At every one of those contacts the leg is already at 0.998 of its own length, so of the 191 mm the pelvis would have to give back, 3.5 mm are within reach. Three ways of taking it anyway were tried and each is priced: one drags the planted foot 872 mm over a stance, one costs a 42.6 mm jump in a single frame, one returns the unpaid part in one place as a reversal.

8. The foot is put down hard. On the last look we approved everything and named one thing: on the last braking step the foot is put down very hard, the whole leg is right, the foot goes through in one jerk at the end. This is the one in the video at about 15 s.

What puts the foot down like that was established rather than guessed. The same "this is a run" amount that owns the limb's shape also reaches the geometry of the support: every braking stance is handed the run's own support vector, 0.99801 of the leg's length at contact on all four braking steps, where the accepted walk lands at 0.99763. The difference is 0.3 mm of leg. A leg at its own length has no way to move its ankle except by moving the hip, so whatever the pelvis does in the 21 ms flight arrives at the ankle instead, and in that flight the pelvis rises. So the last two braking feet arrive going up, +0.07 and +1.03 m/s, the ankle 18 mm higher than one screen frame earlier, and stop dead in a millisecond. Every accepted gait arrives going down: the walk at −0.14 m/s, the run at −0.23 m/s. On the run-to-walk route the same foot arrives travelling backward at −3.85 m/s.

Three answers were built and each was sampled. Tapering the support geometry from the run's to the walk's puts the foot down properly, −0.44 and −1.37 m/s at contact, and costs a vertical acceleration ceiling on the run-to-walk route, 61.1 against the ceiling's 52.6. Placing the feet where a pelvis on the plan would need them costs two contact-solver failures at two run-to-walk contacts. Tapering the pelvis's swing about those contacts makes every motion figure green and costs a hip jerk ceiling by one percent and a knee ceiling by six. Every one of those ceilings is a number I set earlier as the price of continuity. I read all three and moved the work instead of the numbers: the foot is transferred to the next substrate as a rule for the planner, in my words: feet are placed where the pelvis is on the plan, the support geometry tapers from the run's to the walk's, and the foot arrives at contact going down and at the walk's own speed. It is not achieved, and the check that measures every foot's arrival in every clip now names those two feet, so a closed stage cannot be mistaken for a fixed foot.

The lesson worth stating from all eight: per-step checks passed while a whole-window scan failed, and a released bound that turns out not to be the binding one is worth more written down than a ceiling spent quietly to look finished.

How it was judged

The two of us, in real time, two questions per item, ten items, one packet per look: does it respond the way I expect, and does it look natural. Yes or no, and what you saw. A refusal from either of us on any item means no acceptance. Seven packets were shot at the same ten items; six formal looks refused three motions, then two, then one, one, one, and the sixth accepted all ten with the one named condition above. Forty cells, forty yes.

The closure was written down before the first look was taken: what a pass looks like, what a pass with a condition looks like, what a refusal does. So the recording could not be shaped by the verdict. A machine-green check never accepts anything here; it only refuses.

What is not solved

• A stop during a start in one step, and a tap of the walk command giving exactly one step: transferred to the runtime step planner, not met.

• The landing foot on the exit's last two braking steps arrives rising: transferred, not met, named in the check.

• A run re-asserted in the middle of a run exit: the exit plays out first; deferred to a real-time planner.

• The supporting foot slides about 65 mm backward over a braking stance. It is a property of the accepted run, not of the exit, and it was never touched in this stage.

• One jump is left anywhere in the product, in the start's left knee, and the check names it.

• Desktop stills are not deterministic between captures on one machine; the mobile capture target is the instrument.

• No physics, no terrain, no living idle, one actor, one body.

Claim level

Kinematic. Geometry and determinism, nothing physical. No number in this stage became a measurement; every value is a sampled property of compiled engineering motion at one preset. Nothing has run on hardware and nothing has been physically validated.

Feedback

The two questions, on the video: does it respond the way you expect, and does it look natural. Yes or no, and what you saw.