NVIDIA broadened the scope of its graphics research at SIGGRAPH 2026, anchored by 21 accepted technical papers. At the conference, running July 19–23 in Los Angeles, the company put front and center a chain of technologies moving from making footage look realistic, to assembling virtual worlds, to moving bodies within them, to handing those motions off to physical robots. At the center of this is MotionBricks, which generates movement for both in-game characters and the Unitree G1 humanoid from the same neural model. The 21 papers don't come together into a single unified product. Even so, the three layers needed for physical AI—environment, physics, and skill—are starting to take shape, backed by measurements and public code.
This direction isn't new this year. NVIDIA had already connected neural graphics and synthetic data to physics simulation and reinforcement learning through more than 12 papers at SIGGRAPH 2025. What 2026 adds are concrete performance numbers and deliverables: over 350,000 motion clips, integration into the Newton physics engine, real-robot demonstrations, and reusable training pipelines.
MotionBricks: Calling Up 350,000 Motions in 2 Milliseconds
MotionBricks handles more than 350,000 motion capture clips within a single neural foundation model, and the research team's measurements show a throughput of 15,000 frames per second and a latency of 2 milliseconds. Throughput represents parallel processing volume and latency represents single-response time, so the two numbers can't be directly compared as the same performance metric. Still, the design clearly targets a response speed fast enough to place a generative model inside a game engine's control loop.
MotionBricks combines "Smart Locomotion," which takes speed, direction, and style as input, with "Smart Objects," which fills in the motion from approach through contact and beyond—both built on a shared model. In a research-page demo for Unreal Engine 5, the system composed object interactions without wiring an animation graph. According to the research team, it can be applied zero-shot to new downstream tasks without task-specific tagging or additional training. The very same model that animates on-screen characters was also used to drive the Unitree G1. This points to the possibility that motion generation for animation and motion commands for robotics could share the same underlying data asset.
That said, what's public so far is still an early preview. The current code release includes a lightweight G1 demo and a synthetic training pipeline, but the model fully integrated with GR00T Whole-Body Control and the complete training pipeline are targeted for release in about a month. How well third parties can reproduce training at the scale of 350,000 clips will only become clear once that full version ships.
GPC Turns Motion Into Reusable Skills
GPC (Generative Pretrained Controller) bridges the gap between a model that generates motion and a controller that stably drives a physical body. It works in three stages: first, Finite Scalar Quantization converts continuous movement into discrete skill tokens; then, a GPT-style Transformer decoder learns the distribution over token sequences; and finally, Conditional Low-rank Adaptation (CoLA) transfers the pretrained skills to downstream tasks such as parkour.
According to the research page, the tracking controller used for the initial skill quantization was scaled up to a 600-hour dataset, achieving a 99.98% success rate. This figure represents how well the discrete representation could track the original motion—it is not a robot's success rate in unknown real-world environments. CoLA keeps the additional parameters needed for downstream adaptation to under 1%. Rather than retraining a large motion model for each task, the design preserves shared skills while adapting to new use cases with a small number of added parameters.
While MotionBricks focuses on "quickly generating the next motion in response to an instruction," GPC leans toward pretraining a large number of skills and repurposing them for physical control tasks. Additionally, ARDY accepts not only online text instructions but also paths, full-body keyframes, and sparse joint positions and rotations over long time horizons. ARDY, too, was combined with the SONIC tracking policy in a demonstration driving the Unitree G1 in real time. Rather than competing as a single monolithic model, these three research efforts divide up distinct challenges: generation speed, skill transfer, and long-horizon instruction following.
Separate Studies Address 3D Capture, Materials, and Physics
Even with motion ready to hand off to a robot, if the shape, material, and dynamics of the virtual world being trained on are flawed, that motion can't transfer to real hardware. This set of papers also places concrete building blocks on the environment side. ArtiFixer uses a generative model to complete incomplete 3D Gaussian Splatting reconstructions, VideoNeuMat extracts reusable neural materials from the material representations embedded in video generation models, and Mixed Material Point Methods computes the behavior of sand, snow, and elastic solids within the Newton physics engine.
| Research | Gap Addressed | Confirmed Scale/Performance | Remaining Limitation |
|---|---|---|---|
| ArtiFixer | 3D scenes with sparse observation | Generates hundreds of frames in one pass, outperforming existing methods by 1–3 dB PSNR | Unobserved regions are plausible generations, not measurements |
| VideoNeuMat | Relightable 3D materials | Predicts material parameters in a single inference pass from 17 generated frames | The 17 input frames are generated, not real-world captures |
| Mixed MPM | Physics of sand, snow, and elastic bodies | 49 million particles on a single GPU at 4 seconds per frame | Large-scale examples are not real-time |
| ARDY | Long-horizon motion instructions | Sequentially processes text, paths, keyframes, and joint constraints | On real hardware, must be combined with a separate tracking policy |
What this comparison shows is that producing photorealistic imagery alone doesn't constitute a training environment for physical AI. The regions ArtiFixer fills in aren't necessarily measurements of shapes that actually existed in reality. The materials VideoNeuMat extracts don't guarantee the physical properties themselves either. Using a generated environment for robot training requires, beyond visual consistency, a separate verification process for collision geometry, mass, friction, and similar properties.
Mixed MPM extends the range of materials Newton can handle to include sand, snow, and elastic solids. The solver operates as a first-class Newton module, bidirectionally coupled with rigid-body computation, and the demo showed a case where particulate terrain pushes back against a robot, causing the robot to adjust its stride. On the other hand, the large-scale example of 49 million grains of sand flowing into a city model takes 4 seconds per frame on a single GPU. What the paper demonstrates is an expansion of handleable materials and coupling methods—not a result showing instant computation at every scale.
Three Verification Points That Will Determine Whether This Becomes a Practical Foundation
The first verification point is reproducibility. Papers for each study are now readable, and MotionBricks has released initial preview code, but complete versions are not yet all in place. Once released, it will be necessary to confirm whether external developers can reproduce training at the scale of 350,000 clips, and whether 2-millisecond-class response times can be maintained when transferred to different robot morphologies.
The second is whether sim-to-real skill transfer actually succeeds. The Unitree G1 demos showed that generated motion can be handed off to real hardware. However, success rates in environments with different floor friction, joint-to-joint variation, latency, and disturbances can't be determined from the public pages. GPC's 99.98% figure is also a training metric for the tracking controller—it doesn't answer this question.
The third is the correctness of the generated environments. ArtiFixer can make unobserved areas look natural, but there's no guarantee that the shape and material of surfaces a robot would actually touch have been faithfully reconstructed. Once the full MotionBricks release ships, third-party verification is conducted on different real robots, and a process for assigning physical properties to generated scenes is established, these 21 papers will move from being a showcase of graphics research toward becoming a foundation that physical AI developers can actually mix and match for their own use.