99.98% is a tracking rate, and the field is animation

TL;DR. NVIDIA's Generative Pretrained Controllers, presented at SIGGRAPH 2026, tokenise human motion the way a language model tokenises text and generate physical control by next-token prediction. The headline figure is 99.98%, and it is a tracking success rate: how faithfully the quantisation step reproduced motion clips from the 600-hour dataset it was trained on, in physics simulation. It is not a robot's success rate anywhere. Downstream adaptation via Conditional Low-rank Adaptation adds under 1% additional parameters, which is the genuinely useful result. And the structural point matters more than either. Territory 7 found that robot data is scarce because trajectories must be physically performed while text already existed. GPC routes around that by training on human motion capture, which also already existed, and that is a real move rather than a benchmark number.

---

Status: established, and the primary source is open. Shi, Jiang, Tessler and Peng, GPC: Large-Scale Generative Pretraining for Transferable Motor Control, SIGGRAPH Conference Papers '26, arXiv 2606.29148, licensed CC BY-NC-ND. The paper states its own claim precisely. Nothing in this article corrects the authors; it corrects how the figure travels.

---

What the system does

Three stages, and the middle one is the part that resembles a language model.

First, Finite Scalar Quantization converts continuous movement into discrete skill tokens. An encoder maps a sequence of target states to a continuous latent vector, each dimension of which is independently quantised into fixed scalar levels. This defines an implicit codebook without learning an explicit one, which is the simplification over prior vector-quantised approaches.

Second, a GPT-style autoregressive transformer learns the distribution over those token sequences. Once trained, it generates controls for a physically simulated character by predicting the next token.

Third, Conditional Low-rank Adaptation transfers the pretrained skills to downstream tasks, including parkour, while keeping additional parameters under 1%.

The codebook and the control policy are trained jointly, end to end, with reinforcement learning, on a dataset exceeding 600 hours of diverse human behaviour.

Reported emergent behaviours include responsiveness to perturbation and recovery after falling, which were not separately specified as objectives.

What the 99.98% measures

The paper's own wording: a 99.98% success rate in reproducing a vast corpus of motion clips.

Reproducing. From the corpus. In simulation.

It is the tracking success rate of the FSQ controller during the skill-quantisation stage, scaled to the 600-hour dataset. It answers one question: how faithfully can the discrete representation reconstruct motions it was fitted on?

It does not answer how a robot performs in an unknown environment, and one outside analysis states that explicitly.

None of this is a criticism of the authors. The abstract says "reproducing a vast corpus of motion clips" and the paper describes it as a tracking success rate. The number is correctly stated, correctly scoped, and correctly reported.

What happens afterwards is selective transmission: the figure travels, the word "reproducing" and the phrase "tracking success rate" do not, and by two or three hops it reads as a success rate for robot movement.

Why the number is nearly perfect for the actual field

This is a SIGGRAPH paper. The field is physics-based character animation.

In that field, faithfully reproducing a motion capture clip is not a proxy for the task. It is the task. An animation system that reconstructs a captured performance with high fidelity in a physically simulated body has done the thing it was built to do.

So 99.98% is a strong result where it was measured and a weaker one at every step away from there.

Animation: the metric matches the objective almost exactly. Simulated robotics: it establishes that the representation is expressive enough to encode the skills, which is necessary and not sufficient. Physical robotics: it is close to uninformative, because floor friction, joint-to-joint variation, latency and external disturbance are all absent from the thing measured.

That gradient is the general lesson. A benchmark's informativeness decays with distance from the domain it was constructed in, and the decay is invisible because the number does not change.

Where the real advance is

Two things in this paper are more interesting than the headline.

The first is the adaptation result. Under 1% additional parameters to transfer a pretrained controller to a new downstream task, rather than retraining a motion model per task. If that holds, the economics of character control change, because the expensive artefact is built once and specialised cheaply.

The second is structural, and it connects to something Territory 7 established.

The robot learning data article found that roughly a million trajectories existed, 85% of them from four robot platforms, and identified the cause as an asymmetry: text for language models already existed and had to be collected, while robot trajectories do not exist until somebody physically performs them.

GPC does not solve that. It sidesteps it. Human motion capture is a 600-hour corpus that already existed, collected over decades for animation and biomechanics, and it is a source of physically grounded behaviour that nobody had to perform for this purpose.

That is the same move the field made with text, and it is available for human motion specifically because an unrelated industry spent thirty years building the corpus.

Which also bounds it. Motion capture covers what humans did in a studio. It does not cover manipulation of unfamiliar objects, contact-rich assembly, or the long tail of things a robot would need to do in a building, and no comparable pre-existing corpus covers those.

The idea is not new, and the credit belongs somewhere

Framing motion control as next-token prediction predates this paper.

*Radosavovic, Zhang, Shi, Rajasegaran, Kamat, Darrell, Sreenath and Malik published Humanoid locomotion as next token prediction in Advances in Neural Information Processing Systems in 2024.*

GPC's contributions are the joint end-to-end learning of the quantisation and the control policy, the FSQ formulation that avoids an explicit codebook, the scale of the dataset, and the parameter-efficient adaptation. Those are substantial and they are not the same as introducing the framing.

Coverage presenting this as the moment motion became a next-token problem is two years late, which is an ordinary and unimportant error except that it makes the advance sound larger than it is by attributing the whole idea to one paper.

Does it move Territory 7's test

The humanoid deployment article documented seven units in verified deployment, and the concept graph published a falsification test in advance, stated verbatim:

A deployment of more than a hundred units in continuous commercial operation, performing materially different tasks without reconfiguration, at a site the operator did not modify, with modifications disclosed if any, and a published intervention rate per hour.

GPC moves none of it.

It is simulation. The Unitree G1 demonstrations show generated motion handed to real hardware, which is a meaningful step and is a demonstration rather than a deployment. Success rates under varying floor friction, joint variation, latency and disturbance cannot be determined from the public material, and the 99.98% figure does not address them because it was not measuring them.

The test remains unmet and the framework remains unfalsified, which is the correct status and not a triumphant one. A framework that survives because the evidence that would test it does not exist yet has not been confirmed. It has been left alone.

What would move it: an intervention rate per hour on physical hardware, in an unmodified environment, across materially different tasks. That is a disclosure, not a research result, and nothing about GPC makes it more or less likely to be published.

What the related work suggests about latency

A companion NVIDIA system, MotionBricks, is reported at 15,000 frames per second processing speed, which addresses one of the practical obstacles to running a learned controller on hardware.

Latency being solvable is genuinely useful information and it is a different question from robustness. A controller that runs fast enough and falls over in an unfamiliar environment is fast and useless.

The public material establishes the first and leaves the second open, which is worth stating precisely because the two are frequently reported together as though solving one bore on the other.

The gradient, stated as a table

Naming the distances makes the decay legible in a way prose does not.

Four rows, one number, and its informativeness falls to zero across them.

The reason this is worth tabulating rather than asserting is that every step in that table is a legitimate thing somebody might want to know, and the figure is quoted in support of all four. Only the first is supported.

And the table also shows what would fill each row. Generalisation past the corpus requires held-out motions. Control under contact requires simulated perturbation studies, some of which the paper reports as emergent behaviour. Hardware robustness requires trials under varied conditions. Deployment requires an operator to publish. Four different pieces of evidence, none substitutable for another, and the single number stands in for all of them in casual use.

What this says about reading research

The corpus has now examined three cases where a figure was correct at source and misleading at destination, and this is the cleanest of them.

The energy report contained a worked example undercutting the framing its projections carried. The price analysis published a hundredfold range alongside the rate that travelled. Here the paper printed the qualifying word inside the sentence containing the number: reproducing.

One word, in the abstract, in the same clause. It did not survive.

Which suggests the defence is weaker than "read the source" implies. In this case reading the abstract would have been enough, and reading the abstract is what most people quoting a paper believe they have done. The failure is not that the qualification was buried. It is that a number in a sentence is more legible than the sentence.

No process fixes that at the reader's end. The available interventions all sit with whoever writes the second-hand version, and the cheapest one is to quote the verb: not "99.98% success rate" but "99.98% success reproducing training clips". Four extra words, and every downstream hop inherits the scope.

Three things this establishes

A number can be perfectly stated by its authors and still mislead by the time it reaches a reader. The paper says "reproducing a vast corpus of motion clips" and describes a tracking success rate. Neither qualifier survives transmission, and no correction to the paper would fix that because the paper is already correct.

Benchmark informativeness decays with domain distance. 99.98% is a near-perfect metric for physics-based animation, a necessary-but-insufficient signal for simulated robotics, and close to uninformative for physical deployment. The number is identical at all three distances, which is what makes the decay invisible.

And routing around a data constraint is a different achievement from solving it. Human motion capture existed because another industry built it over thirty years. That is available for locomotion and gesture and is not available for contact-rich manipulation, so the move works exactly once and only where a corpus happens to exist.

What it does not establish

That GPC is overhyped by its authors. The paper states its claim accurately, licenses openly, and describes its method in detail. The inflation happened downstream.

That the advance is small. Sub-1% parameter adaptation and joint end-to-end training of quantisation and policy are real contributions, and the emergent perturbation recovery is a genuine result.

That sim-to-real will fail. Nothing here predicts that. The public material simply does not address it, and absence of evidence is being reported as such rather than as evidence of absence.

And nothing about NVIDIA's broader robotics position. This is one paper among twenty-one the company presented at one conference.

What is unresolved

Whether the 1% adaptation figure holds across task types. Parkour is a locomotion task and the reported transfers are within that family.

What the sim-to-real success rate is. Not published, and it is the number that would matter for the robotics claim.

Whether motion capture generalises beyond its collection conditions. Studio capture is a specific environment, and a 600-hour corpus of it is large within that environment and narrow outside it.

And whether an intervention rate ever gets published. Territory 7 found this missing across ten robotics domains, and it is still missing.

The counter-argument

Complaining about how a number travels is not a criticism of the work, and this article spends most of its length on transmission rather than on the paper. The authors did everything correctly. A corpus that repeatedly examines how figures degrade risks mistaking a communication problem for a research problem, and the research here is good.

The animation framing may undersell the robotics relevance. Physics-based simulated control has transferred to hardware before, the Unitree demonstrations exist, and treating a graphics paper as irrelevant to robotics because of the venue is a category error the field itself does not make.

The data-sidestep argument cuts the other way too. If human motion capture is a usable substrate for locomotion, similar pre-existing corpora may exist for other domains that nobody has thought to repurpose. This article treats the move as a one-off and it may be the first instance of a general strategy.

And the falsification test may be set too high. More than a hundred units, unmodified sites, published intervention rates and task variety without reconfiguration is a bar almost no automation technology met early in its deployment, including ones that later succeeded. A test nothing can pass for years is not obviously a good test, and this corpus should say so about its own.

The short version

GPC tokenises human motion and generates physical control by next-token prediction, in three stages: Finite Scalar Quantization into discrete skill tokens, a GPT-style transformer over those sequences, and Conditional Low-rank Adaptation for downstream tasks at under 1% additional parameters. Trained end to end with reinforcement learning on over 600 hours of human behaviour.

The 99.98% is a tracking success rate. It measures how faithfully the quantisation reproduced motion clips from the dataset it was trained on, in physics simulation. The paper says so. It is not a robot's success rate in any environment, and one outside analysis states this explicitly.

And it is nearly the perfect metric for the field it was measured in, because reproducing a captured performance in a simulated body is the job of physics-based animation rather than a proxy for it. The same number is necessary-but-insufficient for simulated robotics and close to uninformative for physical deployment, and it looks identical at all three distances.

The real advance is structural. Territory 7 found robot data scarce because trajectories must be physically performed. GPC sidesteps that using motion capture that already existed, collected over decades by an unrelated industry. That works where such a corpus exists and nowhere else, which excludes contact-rich manipulation.

And the framing is two years old. Humanoid locomotion as next token prediction appeared at NeurIPS in 2024. GPC's contributions are the joint training, the FSQ formulation, the scale and the adaptation, which is plenty without also claiming the idea.

Territory 7's falsification test is unmoved: more than a hundred units, materially different tasks without reconfiguration, an unmodified site, and a published intervention rate per hour. Simulation does not meet it, a demonstration is not a deployment, and a framework that survives because the testing evidence does not exist has been left alone rather than confirmed.

Common questions

What is GPC? Generative Pretrained Controllers, presented by NVIDIA at SIGGRAPH 2026 and published as arXiv 2606.29148 by Shi, Jiang, Tessler and Peng. It converts continuous human motion into discrete skill tokens using Finite Scalar Quantization, trains a GPT-style autoregressive transformer over sequences of those tokens, and generates controls for a physically simulated character by next-token prediction. A third stage, Conditional Low-rank Adaptation, transfers the pretrained controller to downstream tasks while adding under 1% additional parameters.

What does the 99.98% figure actually measure? The tracking success rate of the quantisation controller: how faithfully the discrete representation reproduced motion clips from the 600-hour dataset it was trained on, in physics simulation. The paper's own wording is a success rate in reproducing a vast corpus of motion clips. It is not a success rate for a robot in an unknown environment, and one outside analysis of the work states that explicitly.

Is the paper overstating anything? No. The abstract says reproducing a corpus of motion clips, the method section describes a tracking success rate, and the work is licensed openly with its method described in detail. The inflation happens in transmission, where the qualifying words drop away and a tracking metric reads as a general capability claim. No correction to the paper would fix that, because the paper is already correct.

Why does the field it was measured in matter? Because reproducing a captured performance faithfully in a simulated body is the actual objective of physics-based character animation, not a proxy for it. So the metric matches the goal almost exactly in that domain, drops to necessary-but-insufficient for simulated robotics, and becomes close to uninformative for physical deployment where floor friction, joint variation, latency and disturbance dominate. The figure is identical at all three distances, which is what makes the decay in informativeness hard to see.

What is the genuine advance? Two things. The adaptation result, where under 1% additional parameters transfers a pretrained controller to a new task rather than retraining per task, which changes the economics if it holds. And a structural one: Territory 7 found robot data scarce because trajectories must be physically performed, whereas GPC trains on human motion capture, a corpus that already existed because animation and biomechanics built it over decades. That sidesteps the constraint rather than solving it.

Is next-token prediction for motion a new idea? No. Radosavovic and colleagues published Humanoid locomotion as next token prediction in Advances in Neural Information Processing Systems in 2024. GPC contributes joint end-to-end learning of the quantisation and control policy, the FSQ formulation that avoids an explicit codebook, the dataset scale, and parameter-efficient adaptation. Those are substantial contributions and they are not the same as originating the framing.

Does this change the outlook for humanoid robots? Not on the published evidence. The falsification test this corpus set in Territory 7 asks for more than a hundred units in continuous commercial operation, performing materially different tasks without reconfiguration, at a site the operator did not modify, with a published intervention rate per hour. GPC is simulation work with hardware demonstrations, and success rates under varying friction, joint variation, latency and disturbance are not determinable from the public material. The test remains unmet, which means the framework has been left alone rather than confirmed.

What would actually change the picture? A published intervention rate per hour, on physical hardware, in an environment the operator did not modify, across materially different tasks. That is a disclosure rather than a research result, and nothing about this paper makes it more or less likely to appear. Territory 7 found that figure missing across ten robotics domains and it is still missing.