Humanoid Whole-Body Control
Learned controllers that move a whole humanoid body, balancing, walking and manipulating at once.
Tier Grades
Tier A: Frontier AI labs
feasible
NVIDIA trained SONIC with 21k GPU hours on 700 hours of motion capture.
Tier B: Indonesian AI labs
stretch
21k GPU hours and a humanoid robot are beyond the evidence for Indonesian labs, though the model itself is only up to 42M parameters.
Tier C: Institution units such as ITB
stretch
A 42M-parameter controller is small, but 21k GPU hours and humanoid hardware are not shown for institution units.
Tier D: Organisations such as SEMAI (PoT, PoC, PoV)
not feasible
Building a whole-body controller needs 21k GPU hours and humanoid hardware; delivery work would only use vendor robots.
Where this comes from
Tiers: a) frontier AI labs, b) Indonesian AI labs, c) institution units such as ITB, d) organisations such as SEMAI delivering proof of technology, concept and value. Each rationale cites the anchors below.
Anchors
- ComputeSONIC: 1.2M to 42M parameters trained with 21k GPU hours.
- DataSONIC: over 100M frames from 700 hours of motion capture.
- Team size (proxy)Proxy only (author count of an exemplar paper, not a headcount): 28 authors on the SONIC paper (arXiv 2511.07820).
- Time horizonMulti-year.
- Shelf-life signalsSONIC notes that earlier humanoid controllers were trained on a handful of GPUs; scaling to 21k GPU hours changed the state of the art.
- Pipeline roleTier A trains whole-body controllers on its humanoids; Tier B and Tier C could study released controllers in simulation; Tier D would deploy vendor humanoids rather than build controllers.
Where this comes from
Team size is a proxy (author counts of exemplar papers), not a headcount.
Who and What
Exemplars
Observed actors
Linked models
Links
Papers
- DeepMimic: Example-Guided Deep Reinforcement Learning of Physics-Based Character SkillsXue Bin Peng et al., arXiv preprint, 2018 founding
- GR00T N1: An Open Foundation Model for Generalist Humanoid RobotsNVIDIA et al., arXiv preprint, 2025 state of the art
- Expressive Whole-Body Control for Humanoid RobotsXuxin Cheng et al., arXiv preprint, 2024 state of the art
- Learning Human-to-Humanoid Real-Time Whole-Body TeleoperationTairan He et al., arXiv preprint, 2024 state of the art
- OmniH2O: Universal and Dexterous Human-to-Humanoid Whole-Body Teleoperation and LearningTairan He et al., arXiv preprint, 2024 state of the art
Where this comes from
The papers to read first: where the direction began, its current best, and a survey when one exists.
Sources
- arxiv.org/abs/2511.07820 https://arxiv.org/abs/2511.07820
- GR00T N1: An Open Foundation Model for Generalist Humanoid Robots arxiv.org https://arxiv.org/abs/2503.14734
To cite this page: Project AURION, AI Use-Cases Research & Innovation, "Humanoid Whole-Body Control", https://aurion.semai.id/topics/humanoid-whole-body-control/, records of October 2026.