Chapter 4: GR00T and Humanoid VLAs — System 2 Sees, System 1 Acts

Overview

GR00T is best read not as a single humanoid product, but as NVIDIA's attempt to bind language, vision, and action into a manufacturing skill executor. Since GR00T N1 in 2025, the recent GEAR line has expanded toward N1.7 model cards, a human-to-humanoid motion supply chain around SOMA, BONES-SEED, GENMO, Kimodo, MotionBricks, and GRAIL, SONIC-style whole-body controller scaling, VIRAL and Doorman pixel-to-action sim-to-real, and PLD-style self-improving VLAs ^[1]; ^[27]; ^[21]; ^[22]; ^[23]; ^[24]; ^[25]; ^[26]; ^[17]; ^[18]; ^[19]; ^[20].

The point is not the humanoid demo. A manufacturer should care less about whether a model understands a natural-language instruction in isolation and more about whether that instruction can be grounded in fixtures, tolerances, quality rules, safety interlocks, and human approval paths.

After reading this chapter, you will be able to... - Explain GR00T-style dual-system control as a manufacturing skill executor. - Separate the human-motion dataset, retargeting, controllable reference-motion, and whole-body-control layers below GR00T. - Compare OpenVLA, pi0, FAST, GR00T N1.7, and PLD in terms of action representation, post-training, and deployment cost. - Treat the reference humanoid as a standardized experimental platform rather than production proof. - Identify the data and evaluation harness a manufacturer must own in the first pilot cell.

4.1 GR00T Is an Execution Interface, Not Just a Humanoid

The key design choice in GR00T N1 is the split between high-level reasoning and low-level action generation. System 2 interprets the scene, language instruction, and goal state; System 1 emits short-horizon continuous action chunks ^[1]. In a manufacturing cell, System 2 reads an instruction such as "pick the connector from the left tray, align it with the fixture, insert it, and inspect the latch," while System 1 executes grasping, approach, contact, insertion, withdrawal, and retry behaviors.

That separation fits factory governance. A quality team can inspect whether System 2 respects operating procedures, forbidden states, and inspection rules. A robotics team can separately measure whether System 1 satisfies cycle time, contact stability, collision margin, and recovery requirements. This is easier to approve and roll back than a monolithic end-to-end model with no operational boundary.

The dual-system label, however, is not evidence of production readiness. The GR00T N1.7 model card connects a Cosmos-Reason2-style encoder to a flow-matching action transformer and frames post-training as the path to specific robots and tasks ^[27]. That is a useful starting point, not a factory acceptance test. A factory still needs SKU-specific tests for fixtures, lighting, surface finish, tolerance stack-ups, and operator intervention.

Read as a 2026 NVIDIA stack, GR00T is the execution interface at the top of a six-layer flywheel rather than the whole system. SOMA, BONES-SEED, and GENMO standardize human motion; Kimodo and MotionBricks generate controllable reference motion; SONIC and GEAR-SONIC lower that motion into humanoid whole-body control; and GRAIL synthesizes task-level loco-manipulation demonstrations from 3D assets and video priors ^[21]; ^[22]; ^[23]; ^[24]; ^[25]; ^[26]. Cosmos, Isaac, and Newton then connect world/action data, simulation, and physics validation behind it.

4.2 Reading OpenVLA, pi0, FAST, and PLD Together

OpenVLA made VLA experimentation more accessible by pairing large-scale real-robot demonstrations with an open fine-tuning recipe ^[3]. The Open X-Embodiment and RT-X line showed how data from different robots and tasks can be pooled into one policy family ^[8]. For manufacturers, this raises a practical question: how much cell-specific data is needed before an existing VLA can be adapted to a real process?

pi0 and pi0.5 push toward broader open-world manipulation and long-horizon generalization ^[9]. Their examples are closer to daily environments than factories, but the underlying problem resembles manufacturing manual work: instructions are long, objects vary, and mid-task recovery matters.

FAST addresses a lower-level bottleneck: making action representation shorter and smoother ^[5]. In factory cells, autoregressive decoding cost and high-frequency hand motion translate into real latency and edge-compute constraints. An action tokenizer may look like a paper detail, but it becomes an operating variable for edge deployment and tactile manipulation.

GEAR's PLD work points to a different bottleneck. If a large VLA improves only through expensive human demonstrations, failure-region data remains scarce. PLD trains residual RL actors to probe where the generalist fails, collects rollout data closer to the deployment distribution, and distills that data back into the generalist VLA ^[20]. In manufacturing terms, this asks whether failures merely become new demo requests or whether safe residual exploration can produce recovery data that is absorbed back into the approved policy.

4.3 Whole-Body Controller Scaling: SONIC, GRAIL, VIRAL, Doorman

If GR00T is the high-level policy interface, SONIC, GRAIL, VIRAL, and Doorman describe the lower execution layer that actually moves a humanoid body. SONIC treats motion tracking as a scalable supervision task and uses 700+ hours of motion-capture data, 100M+ frames, and 21k GPU-hours to build a natural whole-body motion prior ^[17]. For manufacturing, the key point is not that a humanoid can move naturally; it is that VLA, VR teleoperation, and kinematic planning can share one token/control interface.

GRAIL extends that execution layer into task-level demonstration generation. Its pipeline combines 3D assets and video foundation model priors to synthesize more than 20,000 loco-manipulation sequences across pick-up, whole-body manipulation, sitting, and terrain traversal, then reports policies trained only on generated data reaching 84% real-world object pick-up and 90% stair-climbing on Unitree G1 ^[26]. This is not factory deployment proof; it is evidence that metric assets and video priors can generate humanoid task-demonstration candidates before physical trials.

VIRAL and Doorman push the pixel-to-action sim-to-real frontier. VIRAL combines a privileged RL teacher, RGB student distillation, tiled rendering, and visual domain randomization to deploy zero-shot continuous loco-manipulation on Unitree G1 ^[18]. Doorman narrows the benchmark to articulated-object interaction and uses staged resets plus GRPO-style fine-tuning to train a pure-RGB humanoid policy ^[19]. These are not production-deployment proofs. They are research evidence for how far Isaac-scale simulation can push body policies before real-cell validation.

4.4 What the Reference Humanoid Means

The reference humanoid should not be read as a claim that one body shape will become the factory default. It is better understood as a standardization device: body, hands, onboard compute, Isaac workflow, and policy interface are packaged so researchers and partners can experiment from a common baseline. The inclusion of tactile hands and Jetson-class edge compute signals that NVIDIA sees the bottleneck not only in seeing, but also in grasping, sensing contact, and recovering after contact.

The 2026 Isaac GR00T Reference Humanoid announcement belongs in that same category. NVIDIA describes a research reference design combining a Unitree H2 Plus body, Sharpa tactile five-finger hands, Jetson AGX Thor T5000, Isaac Teleop/Sim/Lab/ROS, and the GR00T software stack ^[27]. This is not a factory-ready humanoid. It is an integration artifact for experimenting with data capture, simulation, policy learning, and edge deployment from a shared baseline.

Manufacturing automation still has to separate tasks that need a humanoid form from tasks that do not. In a brownfield cell that must share human spaces and tools, a humanoid or bimanual mobile manipulator may be useful. In a cell that can use fixed fixtures, dedicated end-effectors, and simple part feeding, a humanoid can add cost and safety burden without improving the process.

So the reference humanoid is most useful as a question generator. Does the process need a human-like reach envelope? Is bimanual coordination the actual bottleneck? Is a tactile hand required, or would a compliant gripper plus a force-torque sensor be enough? Where can a human operator intervene safely when the model fails?

4.5 Designing Manufacturing Skill Executors

The unit of factory deployment is not an impressive robot behavior. It is an approvable skill. A skill has start conditions, input state, allowed tools, action bounds, quality judgment, failure codes, and recovery paths. GR00T and VLA systems should be interpreted as execution layers that call and compose these skills.

For example, a "cable insertion" skill does not end with recognizing the part. It includes connector orientation, pre-insertion alignment, contact detection, insertion-force profile, latch confirmation, retry behavior, and final image inspection. System 2 manages procedure and state; System 1 produces smooth contact-rich actions. A separate vision or force/tactile check may own the quality decision.

This structure makes model replacement easier. OpenVLA, GR00T, pi0, or an internal policy can change while the task schema and evaluation harness remain stable. The asset a manufacturer must own is not one checkpoint. It is task data, simulation assets, failure taxonomy, and acceptance tests.

4.6 Manufacturing Cell Checkpoint

The first pilot should not ask "can a humanoid do this?" It should ask "can this cell's skill be versioned and evaluated?"

4.7 What To Learn Next

This chapter interpreted GR00T and VLAs as manufacturing skill executors, then used SONIC, VIRAL, Doorman, and PLD to show how that execution layer is expanding into whole-body control, sim-to-real, and self-improvement. The next chapter moves one level up to Jim Fan and GEAR's research map. MineDojo, Voyager, VIMA, Eureka, DreamDojo, DreamZero, and EgoScale are not factory-ready robot products; they are experiments in growing open-ended agents through skill libraries, reward automation, world models, and human-video data flywheels.

References

1. Johan Bjorck et al. (2025). GR00T N1: An Open Foundation Model for Generalist Humanoid Robots. arXiv preprint. https://arxiv.org/abs/2503.14734
2. Ankur Handa et al. (2023). DeXtreme: Transfer of Agile In-hand Manipulation from Simulation to Reality. ICRA 2023. https://arxiv.org/abs/2210.13702
3. Moo Jin Kim et al. (2024). OpenVLA: An Open-Source Vision-Language-Action Model. arXiv preprint. https://arxiv.org/abs/2406.09246
4. Kadalagere Sampath et al. (2023). Review on Human-like Robot Manipulation Using Dexterous Hands. Cognitive Computation and Systems (IET/Wiley). https://doi.org/10.1049/ccs2.12073
5. Karl Pertsch et al. (2025). FAST: Efficient Action Tokenization for Vision-Language-Action Models. arXiv preprint. https://arxiv.org/abs/2501.09747
6. Uikyum Kim et al. (2021). Integrated Linkage-Driven Dexterous Anthropomorphic Robotic Hand. Nature Communications. https://doi.org/10.1038/s41467-021-27261-0
7. Junjie Wen et al. (2024). TinyVLA: Towards Fast, Data-Efficient Vision-Language-Action Models for Robotic Manipulation. arXiv preprint. https://arxiv.org/abs/2409.12514
8. Open X-Embodiment Collaboration (2024). Open X-Embodiment: Robotic Learning Datasets and RT-X Models. ICRA 2024. https://arxiv.org/abs/2310.08864
9. Physical Intelligence (2025). pi0.5: A Vision-Language-Action Model with Open-World Generalization. arXiv preprint. https://arxiv.org/abs/2504.16054
10. Karl Pertsch et al. (2024). Towards Generalist Robot Policies: What Matters in Building Vision-Language-Action Models. arXiv preprint. https://arxiv.org/abs/2412.14058
11. Zhenyu Wei et al. (2024). D(R,O) Grasp: A Unified Representation of Robot and Object Interaction for Cross-Embodiment Dexterous Grasping. IEEE International Conference on Robotics and Automation (ICRA). https://arxiv.org/abs/2410.01702
12. Antonio Bicchi (2000). Hands for Dexterous Manipulation and Robust Grasping: A Difficult Road Toward Simplicity. IEEE Transactions on Robotics and Automation. https://doi.org/10.1109/70.897777
13. Alexander Khazatsky et al. (2024). DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset. RSS. https://arxiv.org/abs/2403.12945
14. Kenneth Shaw et al. (2023). LEAP Hand: Low-Cost, Efficient, and Anthropomorphic Hand for Robot Learning. Robotics: Science and Systems (RSS) 2023. https://arxiv.org/abs/2309.06440
15. NVIDIA (2026). GR00T-N1.7-3B Model Card. Hugging Face. https://huggingface.co/nvidia/GR00T-N1.7-3B
16. NVIDIA (2026). NVIDIA Announces NVIDIA Isaac GR00T Reference Humanoid Robot for Academic Research. NVIDIA Investor Relations. https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-Announces-NVIDIA-Isaac-GR00T-Reference-Humanoid-Robot-for-Academic-Research/default.aspx
17. Zhengyi Luo et al. (2026). SONIC: Supersizing Motion Tracking for Natural Humanoid Whole-Body Control. arXiv preprint. https://arxiv.org/abs/2511.07820
18. Tairan He et al. (2025). VIRAL: Visual Sim-to-Real at Scale for Humanoid Loco-Manipulation. arXiv preprint. https://arxiv.org/abs/2511.15200
19. Haoru Xue et al. (2025). Opening the Sim-to-Real Door for Humanoid Pixel-to-Action Policy Transfer. arXiv preprint. https://arxiv.org/abs/2512.01061
20. Wenli Xiao et al. (2025). Self-Improving Vision-Language-Action Models with Data Generation via Residual RL. arXiv preprint. https://arxiv.org/abs/2511.00091
21. Jiefeng Li et al. (2025). GENMO: A GENeralist Model for Human MOtion. arXiv preprint. https://arxiv.org/abs/2505.01425
22. NVIDIA SOMA team (2026). SOMA: Unifying Parametric Human Body Models. arXiv preprint. https://arxiv.org/abs/2603.16858
23. BONES Studio (2026). BONES-SEED: Skeletal Everyday Embodiment Dataset. Hugging Face dataset. https://huggingface.co/datasets/bones-studio/seed
24. Davis Rempe et al. (2026). Kimodo: Scaling Controllable Human Motion Generation. arXiv preprint. https://arxiv.org/abs/2603.15546
25. Tingwu Wang et al. (2026). MotionBricks: Scalable Real-Time Motions with Modular Latent Generative Model and Smart Primitives. ACM Transactions on Graphics / arXiv. https://arxiv.org/abs/2604.24833
26. Tianyi Xie et al. (2026). GRAIL: Generating Humanoid Loco-Manipulation from 3D Assets and Video Priors. arXiv preprint. https://arxiv.org/abs/2606.05160
27. NVIDIA (2026). SOMA Retargeter. GitHub repository. https://github.com/NVIDIA/soma-retargeter