GPT-6 Astra aces RoboDojo zero-shot, shaking embodied AI startups' faith in bespoke models

量子位 · wechat · 2026-11-08

QbitAI reports that just two months after GPT-6 Astra's release, the embodied AI industry is reeling: some robotics startups are cutting algorithm roles, and teams that insisted on training native embodied foundation models are pivoting to approaches built on general-purpose LLMs.

Chen Tianxing (HKU MMLab PhD, first author of the RoboTwin/RoboDojo benchmarks) says models scoring 20-30 have jumped to 50-60 when combined with Astra-based solutions. Evaluations found Astra exhibits three capabilities that were once standalone research agendas—without task-specific post-training or demonstrations, from language instructions alone: in-context learning (mimicking a human demo video, with online correction when camera views or coordinate frames shift), OOD generalization and error correction (re-grabbing a bumped cup), and open-vocabulary manipulation. The report cautions Astra's real-robot testing was halted over safety issues, so simulation scores don't yet equal real-world productivity.

Founders push back: Xu Huazhe (Brock Robotics) notes complex contact, force control and deformable objects remain hard; Qian Wang (AutoVariable) stresses data, real-world eval and deployment infrastructure don't vanish. High cloud inference latency hampers real-time control, but quantization/distillation/on-device stacks can mitigate it, and control frequency ≠ inference frequency. Chen's verdict: embodied startups should compete on economics—cheap specialized models for standardized assembly lines, general models for open-ended home tasks—and refocus R&D on physical interaction, embodiment control and deployment. The sector mood has turned to "investors pulling back from embodied AI" as the hundred-robot battle enters its knockout stage.

Original post →

More from Embodied

Embodied channel →