Agibot Chief Scientist Rejects LLM Path for Robotics

For practitioners, the core takeaway is that robotics requires production-grade data and deployment pipelines, not just larger models. Embodied systems combine sensors, control loops, environment variability, and long-tail failure modes; without standardized data schemas and real-world feedback loops, gains from model scaling will be limited.

Reporting by Kr-Asia covers comments from Luo Jianlan, associate professor at the Shanghai Innovation Institute and chief scientist at Agibot, who argues that robotics will not get a "GPT moment" by simply following the LLM development path. Kr-Asia describes a sector shift in China over the past six months, away from a fixation on robot degrees of freedom toward attention on data, models, infrastructure, and their interaction in deployment. The piece notes Luo's academic and industry background at UC Berkeley, Google X, and DeepMind, and reports his assessment that many "embodied foundation models" are effectively mid-training or fine-tuned models because of insufficient high-quality real-robot interaction data.

Kr-Asia's reporting highlights several technical bottlenecks that commonly recur in embodied-AI discussions: scarcity of multi-robot, multi-task datasets that include corrective interventions and failure cases; heterogeneity of robot embodiments that complicates transfer learning; and the persistent sim-to-real gap for edge-case behaviors. These are industry-wide constraints, not claims about any single company's internal roadmap.

Observers and practitioners should track three indicators that would support the sort of system-level loop Luo describes:

• •emergence of community data standards or shared benchmarks for multi-body interaction data
• •investments in long-duration, safety-conscious field deployments that produce corrective feedback data
• •tooling for cross-robot transfer and evaluation that measures performance on rare failure modes

Companies and research groups aiming to scale embodied AI typically pivot from isolated model improvements toward integrated data collection and deployment engineering once they hit diminishing returns from model-only gains. If public engineering efforts and datasets follow, progress may accelerate; if not, model scaling alone is unlikely to deliver general-purpose robotic competence. Reporting by Kr-Asia provides the primary source for Luo's views; Agibot has not been quoted verbatim in the coverage cited.