GigaBrain-0.7: Scaling Embodied Foundation Models to Emergent Capabilities with a Three-System Architecture
GigaBrain Team, Angen Ye, Axiang Sun, Can Jin, Chenxi Cheng, Chong Shi, Dengke Shang, Dingqian Zhang, Guan Huang, Guangqiang Wang, Guangqing Ding, Guo Li, Hangcong Li, Hengyu Zhong, Hongtao Lu, Jianbo Qin, Jiming Mao, Jing Zhu, Jindi Lv, Jingzhi Cui, Junjie Xie, Junyi Bao, Kai Liu, Lei Yuan, Limin Long, Lv Feng, Mingming Yu, Peng Li, Pengfei Yi, Qi Li, Qianli Zhang, Qingfang Li, Qitang Hu, Rui Zhang, Shaoyan Sun, Shibo Sun, Shiying Duan, Tenghui Chen, Tianze Liu, Weijie Ke, Wenyao Xue, Xiaofeng Wang, Xiaoyu Tian, Xinyu Liu, Xinze Chen, Yang Wang, Yankai Wang, Yejun Zeng, Yifan Li, Yifei Nie, Yilong Li, Yilong Liu, Yongchao Feng, Yumeng Wang, Yun Ye, Zhichao Liu, Ziheng He, Zonghai Yang, Zheng Zhu
cs.RO
2026-08-17
GigaBrain-0.7 is a three-system VLA pretrained on 37k hours across 16 robots; post-training reaches 74.1% on Maker H01 complex tasks vs 45.2% for π0.5.
Most vision-language-action (VLA) models graft a pretrained vision-language model onto an action head and map the current observation plus an instruction into motor commands. That recipe already works for short tabletop skills. It starts to fail once the data mix spans many robot bodies, or the task runs long enough that the same-looking scene means different things at different stages.
Heterogeneous logs disagree on degrees of freedom, coordinates, and control rates, so naive mixing can cancel rather than transfer. Architecturally, most VLAs stay reactive: they do not predict a future scene, and they do not score whether the last chunk actually advanced the task. Understanding, prediction, action, and post-deployment correction usually live in separate training stages. GigaBrain-0.7 tries to run those four jobs as one lifecycle.
Three systems, with explicit interfaces.
The real bulk is data. After cleaning, the embodied corpus is 37,257 hours across 16 robot types: 20,536 hours of real robots (55%), 8,252 hours of UMI near-hand demos, 2,862 hours of egocentric human video, 1,454 hours of simulation, and 4,153 hours of world-model generation, plus 272 million vision-language samples. In-house Maker H01 accounts for 43.8% of real-robot frames, Maker M01 another 20.1%. Everything is converted to LeRobot v3.0, with states and actions aligned as left arm, right arm, head, waist, base, and missing axes masked rather than zero-filled.
Pretraining is one-stage: next-token loss on language, subtasks, and discrete actions, plus flow matching on continuous chunks. Flow-matching gradients are scaled by a factor below 1 before they enter the VLM backbone (Soft Knowledge Insulation), so the semantic representation is not overwritten. At post-training, System 3 stays frozen and its subgoal image plus progress bit condition System 1; inference always conditions on positive progress. The same policy is then refined with advantage-weighted regression on its own rollouts, then with human-in-the-loop online RL.
Architecture first. PaliGemma2 (3.5B total) is the only backbone with nonzero shirt-folding success (30%); Gemma 4 at 8.5B scores 0% on that task. Dual-stream coupling hits 50% / 88% / 30% on desk cleaning, fruit picking, and shirt folding. Last-layer and multi-layer cross-attention both fold at 0%.
Zero-shot on real robots already executes, then falls off out of distribution:
| Platform | Language follow ID / OOD | Complex manip. ID / OOD |
| AgileX PiPER | 94.5% / 36.5% | 63.3% / 26.7% |
| Maker H01 | 55.8% / 16.7% | 58.3% / 13.3% |
Button pressing is 0% OOD on both platforms.
After task-specific post-training:
| Method | PiPER language / complex | H01 language / complex |
| π0.5 | 88.8% / 76.6% | 75.2% / 45.2% |
| GigaBrain-0.1 | 76.1% / 64.8% | 69.6% / 43.2% |
| GigaBrain-0.7 | 91.5% / 84.9% | 84.2% / 74.1% |
On the humanoid, complex manipulation opens a nearly 29-point gap over π0.5. The System 3 ablation is cleaner still: gift wrapping goes from 0% (base) to 20% with the subgoal image, 60% with the value bit, and 80% with both. On clothes folding, where success is already 100%, the progress score still rises from 68.3% to 88.3% and completion time drops from 107 s to 75 s.
Simulation: RoboTwin 2.0 Co-Train overall 67.35, with Hard at 67.9 versus 46.0 for π0.5; π0.5 still leads Easy by 3.9 points. EBench success 33.30 versus 28.08 for π0.5. RoboColiseum leads all four axes among reported open models, 0.8166 instruction following and 0.4729 spatial reasoning. Human-in-the-loop RL lifts four real-robot tasks from 30% after SFT to 57.5% offline and 100% online, at roughly 10 to 20 trials per configuration.
Three takeaways for anyone training a generalist robot policy. Heterogeneous data can share one model if states are aligned and Soft KI damps action gradients, instead of blocking them. A reactive VLA is not enough: an explicit future image and a progress bit take gift wrapping from 0% to 80%. Scale bites hardest on deformable objects; clothes folding steepens with data ratio, fruit picking rises more smoothly. The paper calls this emergence. A tighter reading is a coverage threshold for hard skills.
Code and weights are promised. The reusable pieces are the data pipeline (LeRobot v3.0, cross-embodiment masks, LLM instruction rewrite) and the frozen world model used as a conditioner at post-training. On PiPER language following, π0.5 is only 2.7 points behind. The gap shows up on the humanoid and on Hard sim.
The authors report the ID-OOD drop themselves: language following 94.5% to 36.5% on PiPER, 55.8% to 16.7% on H01. When the interaction structure changes, as in OOD button press, zero-shot success is zero. Contact-rich tasks are far from saturated: rice sweeping on PiPER is 40%.
π0.5 still wins RoboTwin Easy. Real-robot hours are heavily in-house. System 3 is frozen at post-training and never jointly updated with the policy. The 100% online-RL number includes human corrections; the intervention protocol is deferred to a later report, and trial counts are small. Bigger backbones are not monotonically better: 8.5B Gemma 4 fails shirt folding. Dual-stream inference costs 0.221 s versus 0.073 s for last-layer cross-attention.
"Emergent" is in the title. What the plots show is a steep data-to-success curve on hard skills, not a new class of capability appearing from nowhere.