SONIC (Humanoid Control)

SONIC, short for Supersizing mOtion tracking for Natural humanoId Control, is a physics-based whole-body motion tracking and control system developed by Nvidia. It was designed and evaluated as a single generalist controller for a Unitree G1 humanoid. The controller maps desired human, robot, or mixed motion into target joint positions while reacting to the robot's measured state. NVIDIA's researchers use the name GEAR-SONIC for the public implementation and checkpoints.^[1][3]

SONIC is a low-level motion controller, not a complete autonomous robot intelligence system. Video, text, music, virtual-reality input, gamepad commands, and vision-language-action policies can all supply reference motions through its common interface, but the perception, language processing, or planning that creates those references happens in separate upstream components. The peer-reviewed study appeared in Science Robotics on August 12, 2026, following an arXiv preprint first posted in November 2025.^[1][2]

Development and publication

The project was produced by a 28-author NVIDIA team. Zhengyi Luo, Ye Yuan, Tingwu Wang, Chenran Li, and Fernando Castañeda were designated co-first authors, while Linxi Jim Fan and Yuke Zhu were equal advisers. The journal records the study in volume 11, issue 117, as article eaed4592. Its disclosure states that NVIDIA funded the research, that the authors were current or former NVIDIA employees at the time of the work, and that the company filed a related patent application.^[1]

The paper frames motion tracking as a way to scale humanoid robot control without designing a separate task reward for every behavior. A motion capture frame supplies a target pose at each time step, so a tracker receives dense supervision across walking, dancing, gestures, tool use, combat motions, and other behavior classes. The authors tested whether increasing the motion corpus, neural-network capacity, and training compute improved tracking and generalization together.^[1][2]

Public materials use GEAR-SONIC as the release name. NVIDIA's repository contains training and deployment code, pretrained checkpoints, simulation evaluation tools, and teleoperation documentation. The code is distributed under Apache License 2.0, while model weights use the NVIDIA Open Model License.^[4][6]

Training data

The source motion capture collection contains about 700 hours. The team retargeted the human recordings to the G1 with General Motion Retargeting and PyRoki, then removed movements considered physically infeasible for the robot, including stair climbing and seated activities. The resulting training set contains 611 hours, more than 100 million frames sampled at 50 Hz, 317,189 clips, and 8,447 motion subcategories within 33 main categories. Those categories include locomotion, gestures, acting and roleplay, combat, object manipulation, dance, injured gait, and tool use.^[1][2]

Two held-out splits test different kinds of generalization. test-content contains 7,016 clips totaling 15 hours from 182 subcategories absent from training. test-repetition contains 9,395 clips totaling 12 hours; its 1,088 subcategories also occur in training, but the individual performances do not. The paper additionally evaluates 68,326 motions from PHUMA, which uses a different source and retargeting pipeline.^[1]

A portion of the source data was released as BONES-SEED. Its dataset card lists 142,220 annotated sequences, 288 hours, and 522 actors, with SOMA and G1 representations, natural-language descriptions, and temporal annotations. The Hugging Face repository is publicly listed but gated, and users must accept its access conditions. BONES-SEED therefore covers less than half of the 611-hour training corpus and is not an unrestricted release of the entire dataset.^[1][5]

Policy architecture and training

SONIC formulates tracking as a Markov decision process and trains the policy with Proximal Policy Optimization in NVIDIA Isaac Lab. Its proprioceptive input includes joint positions and velocities, root angular velocity, gravity direction, the previous action, and a ten-step history. A motion command supplies the desired behavior. The policy outputs target joint positions, which proportional-derivative controllers at individual joints attempt to follow.^[1][2]

The reward combines tracking terms for root and body-link pose and velocity, including the head, wrists, and ankles, with penalties intended to encourage smooth and stable motion. During training, domain randomization varies simulated physical parameters, adds root-velocity perturbations that act like pushes, and perturbs reference commands. These measures support sim-to-real transfer, but do not guarantee stability under every command or contact condition.^[1]

Three multilayer-perceptron encoders accept different command formats. The robot encoder reads future robot joint motion. The human encoder reads future three-dimensional human joint positions. The hybrid encoder combines current head and hand keypoints with a lower-body robot reference. Each encoder maps its input to a shared latent representation, which Finite Scalar Quantization converts into universal motion tokens. A common control decoder combines those tokens with proprioception to produce joint targets. An auxiliary decoder reconstructs robot motion and helps align the three input formats, including learning an implicit human-to-robot mapping.^[1]

Training adds reconstruction, token-consistency, and cycle-consistency objectives to the PPO loss. An adaptive sampler gives additional attention to motions on which the policy fails while retaining coverage of the corpus. The scale experiments use data subsets of 4 million, 10 million, 22 million, and more than 100 million frames; networks with 1.2 million, 16 million, or 42 million parameters; and runs on 16, 32, or 128 GPUs. The largest run used 128 GPUs for about seven days, or roughly 21,000 GPU-hours, and all physical experiments used the 42-million-parameter model.^[1][2]

Planning and control interfaces

A separate kinematic planner converts interactive commands into short reference motions for the tracker. It generates segments lasting 0.8 to 2.4 seconds, can replan every 100 ms or when a command changes, and uses a critically damped spring model to smooth abrupt changes in root position and heading. The paper reports planner inference below 5 ms on a laptop and about 12 ms on a Jetson Orin. Through this layer, the same tracker can execute commanded direction, speed, locomotion style, squatting height, kneeling, crawling, and boxing without retraining.^[1]

For video, text, and music control, a separate model called GEM estimates or generates human motion, which SONIC then tracks. For virtual reality teleoperation, the full-body interface uses a PICO headset, ankle trackers, and handheld controllers. A three-point version uses only the headset and controllers for head and hand poses while the planner supplies lower-body motion.^[1][3]

The researchers also connected Isaac GR00T N1.5, a vision-language-action model, to SONIC. After fine-tuning on teleoperated demonstrations, GR00T produced a 78-dimensional action consisting of a 64-dimensional motion token and 14 hand-joint angles. SONIC decoded the motion token into whole-body control. This division matters: GR00T selected actions from camera and language input, while SONIC supplied fast motion execution.^[1]

All reported physical experiments used a G1 with 29 actuated joints and an onboard Jetson Orin. TensorRT and CUDA Graph acceleration produced a reported policy forward time of 1 to 2 ms. The deployment stack ran policy inference at 50 Hz, command streaming at 500 Hz, operator input at 100 Hz, and kinematic planning at 10 Hz.^[1]

Evaluation

The primary success metric treats a motion as failed if the robot's root height or an end-effector height differs from the reference by more than 0.25 m. This is a local tracking criterion intended to capture falls and large pose failures, not a measure of global route following. MPJPE-L is the root-relative mean per-joint position error over 14 body links. The authors also report velocity and acceleration differences.^[1]

Scroll sideways for more →

Scaling curves were evaluated at six checkpoints per configuration. Welch tests comparing scale endpoints reported P values no greater than 0.011, and the paper reports P below 0.001 for its baseline success comparisons. On the physical robot, the largest tracking gap relative to simulation was at the feet: 53.7 mm versus 29.0 mm. Each of the 124 real-world motions was attempted only once.^[1]

The baseline table requires a major qualification. BeyondMimic, Any2Track, and GMT were trained on different source datasets and retargeting pipelines. The paper describes the result as a comparison of cross-dataset generalization and scaling, not a data-matched experiment that isolates the tracking algorithm. Tracking-error averages also include successfully completed motions rather than representing early-terminated failures in full.^[1]

The GR00T integration was tested on five task groups, with carrot and scrub variants sharing one object-pickup policy.^[1]

Scroll sideways for more →

The study reports a 75% mean after averaging carrot and scrub as one object-pickup task. Outcomes were strict binary judgments with no partial credit. In a three-task action-space ablation, GR00T averaged 68% success when predicting quantized motion tokens and 27% when predicting explicit SMPL poses, a reported difference of 42 percentage points after rounding.^[1]

Release and later use

The official release includes code, documentation, G1 checkpoints, and simulation and deployment utilities. NVIDIA's model card lists a default checkpoint, a low-latency variant with about 80 ms of reference lookahead rather than about 200 ms, and a later v1.1 checkpoint. An archival code snapshot and the tabulated data underlying the paper's figures are deposited on Zenodo under DOI 10.5281/zenodo.21273312.^[4][6][7]

A May 2026 arXiv preprint, Any2Any, used pretrained SONIC models as source policies for adaptation to LimX Oli and Luna humanoids. Its authors report using about 1% of the data and compute of full training. This is evidence of research reuse, not direct cross-platform support in the original release: Any2Any adds kinematic alignment and parameter-efficient dynamics adaptation, and its claim had not undergone journal peer review as of August 2026.^[8]

Limitations

The paper explicitly says it does not formally treat operational safety or energy efficiency during extended deployment. Although command perturbation and smoothing improve robustness, extreme conditions and highly dynamic motions can still make the robot lose balance. The official teleoperation documentation accordingly recommends testing in simulation, maintaining a clear safety zone, and keeping an operator ready to trigger an emergency stop.^[1][4]

The journal study evaluates only one robot morphology. Source motions that could not be retargeted to the G1 were filtered out, so breadth within the curated corpus does not imply the ability to reproduce every human motion or sustained contact configuration. The public dataset is partial and gated. Real-world tracking uses one attempt per motion, while each autonomous task condition uses only 10 or 20 trials. Those experiments demonstrate feasibility but do not establish production reliability or long-duration safety.^[1][5]

Finally, the simulation baseline comparison is not data-matched, physical foot placement is less accurate than in simulation, and the work's funding, employment, and patent disclosures create a direct institutional interest that readers should consider when interpreting the authors' claims.^[1]

References

1. ^Zhengyi Luo, Ye Yuan, Tingwu Wang, Chenran Li, Fernando Castañeda, et al. *SONIC: Supersizing motion tracking for natural humanoid whole-body control*. Science Robotics 11, no. 117, eaed4592. August 12, 2026. doi.org/...scirobotics.aed4592
2. ^Zhengyi Luo et al. *SONIC: Supersizing Motion Tracking for Natural Humanoid Whole-Body Control*. arXiv:2511.07820v4. August 13, 2026. arxiv.org/...2511.07820v4
3. ^NVIDIA Research. *SONIC: Supersizing Motion Tracking for Natural Humanoid Whole-Body Control*. Project page. 2026. nvlabs.github.io/GEAR-SONIC
4. ^NVlabs. *GR00T-WholeBodyControl*. Source code and documentation. GitHub. Accessed August 14, 2026. github.com/...GR00T-WholeBodyControl
5. ^BONES Studio. *BONES-SEED*. Dataset card. Hugging Face. Accessed August 14, 2026. huggingface.co/...seed
6. ^NVIDIA. *GEAR-SONIC*. Model card and checkpoints. Hugging Face. Accessed August 14, 2026. huggingface.co/...GEAR-SONIC
7. ^Zhengyi Luo et al. *SONIC: Supersizing Motion Tracking for Natural Humanoid Whole-Body Control*. Zenodo software documentation record. July 9, 2026. doi.org/...zenodo.21273312
8. ^Ming Yang, Tao Yu, Feng Li, and Hua Chen. *Any2Any: Efficient Cross-Embodiment Transfer for Humanoid Whole-Body Tracking*. arXiv:2605.23733. May 22, 2026. arxiv.org/...2605.23733