全景 · GR00T Whole-Body Control
NVIDIA 人形动作技术栈，一块一块拆
MotionBricks 不是孤立存在的，这正是报道总把它写拧的原因。GR00T Whole-Body Control 是一组分层——生成、控制、数据、重定向——每层都有自己的仓库或页面。下面是地图，每个盒子都链接到 NVIDIA 自己描述它的地方。
各层
- MotionBricks —— 动作生成。项目页上描述的实时运行时；README 称它与 GEAR-SONIC 一起，是“锚定”GR00T WBC 计划的两个项目之一——从指令生成动作帧的那一层。
- GEAR-SONIC —— 全身控制。控制器层，见其项目页，与 MotionBricks 同在 GR00T-WBC 单仓库（GitHub）：消费参考动作并以 50 Hz 在机器人上执行。三个 checkpoint 的细节见下表。
- GR00T N1.5 / N1.6 decoupled WBC。GR00T N 系列模型线与同一生态里的解耦全身控制工作搭配；本帖只点名位置——当前公开内容以 GR00T-WBC 仓库为准，别信第三方转述。
- BONES-SEED —— 数据。35 万段剪辑声明背后的训练语料：README 指向 bones.studio/datasets，另有 Hugging Face 数据集页。下载实操见这篇。
- SOMA Retargeter —— 转换管线。基于 Newton 的求解器，把 SOMA 动捕重定向到 Unitree G1，产出 MotionBricks 的训练数据（GitHub）。
- Kimodo —— 离线兄弟项目。按 README Related Work 一节，“focused on offline motion generation, complementary to MotionBricks' real-time runtime”（项目页、仓库）。
SONIC 的三个 checkpoint，直接摘自模型卡
All three models use the SONIC universal-token controller, produce 64-dimensional latent motion tokens, run the controller at 50 Hz, and support SMPL pose, G1 motion reference, and teleoperation inputs.
| Checkpoint | 参考 lookahead | 设计用途 |
|---|---|---|
| Default SONIC (original release) | 10 future frames at 20 ms spacing, approximately 200 ms of reference lookahead | Default general-purpose SONIC controller for motion tracking, planning, teleoperation, and compatibility with existing deployments. G1 and teleoperation future-reference observations use step5. |
| Low-latency teleoperation · 文件 | 4 future frames at 20 ms spacing, approximately 80 ms of reference lookahead | Intended for more responsive whole-body teleoperation and VLA execution. G1 and teleoperation future-reference observations use step1. Use its encoder, decoder, and observation config together. |
| SONIC v1.1 · 文件 | 10 future frames at 20 ms spacing, approximately 200 ms of reference lookahead | Uses robot-heading-normalized target orientation and was trained with wrist-pose augmentation. Intended for heading-stable whole-body teleoperation and SONIC-backed VLA policies that use this controller. G1 and teleoperation future-reference observations use step5; this is not the low-latency model. |
The lookahead values describe the reference horizon presented to the controller. They are not measurements of total end-to-end teleoperation latency, which also includes sensing, networking, preprocessing, and inference.
发布时间线
| 日期 | 发布内容 |
|---|---|
| 2026-04-27 | MotionBricks preview release — interactive G1 demo, pretrained checkpoints (VQVAE · pose · root), synthetic training code, motion-representation docs, GIF gallery. |
| 2026-05-07 | GR00T-WBC: end-to-end VLA workflow on G1 (teleop data → Isaac-GR00T N1.7 fine-tune → SONIC deployment). |
| 2026-06-16 | GR00T-WBC: low-latency teleoperation checkpoint (4-frame lookahead) + Isaac Teleop Setup docs. |
| 2026-07-23 | GR00T-WBC: SONIC v1.1 checkpoint (robot-heading-normalized, wrist-pose augmentation). |
值得注意的模式：日期交错出现。这段时间窗口里 MotionBricks 的预览（2026-04-27）最先落地；WBC 仓库随后回应以 VLA 工作流、低延迟 checkpoint 与 v1.1。技术栈整体推进——这正是“只有生成层开不了机器人”的原因。想单看运行时这一块，见安装指南；各层变动时，本站索引的指南会跟进。