Agibot Chief Scientist Rejects LLM Path for Robotics
Luo Jianlan, chief scientist at Chinese humanoid robotics company Agibot and associate professor at the Shanghai Innovation Institute, argues embodied intelligence cannot simply copy the LLM scaling-law playbook, according to a June 29, 2026 interview published by KrASIA (originally reported by 36Kr). Luo says many so-called embodied foundation models are closer to mid-training or fine-tuning than true pretraining, because high-quality multi-scenario robot interaction data -- including failures, corrections, and long-tail events -- remains scarce; a declining training loss, unlike in LLMs, does not reliably predict real-world deployment success for robots. The interview traces a six-month shift in China's robotics sector away from fixation on robot degrees of freedom toward system-level concerns: whether data, models, and infrastructure can reinforce each other in a real-world deployment loop. Agibot's concurrent open-sourcing of the AGIBOT WORLD 2026 dataset, over 1 million trajectories reported across prior releases, offers a concrete example of the data-infrastructure investment Luo describes.
For practitioners, the useful signal here is not that a robotics executive is skeptical of LLM-style scaling -- it's the specific diagnosis of why: real-world interaction data, not parameter count, is the binding constraint, and that reframes what "progress" should look like in embodied AI over the next year.
What happened
In a June 29, 2026 interview published by KrASIA (originally reported by 36Kr), Luo Jianlan, associate professor at the Shanghai Innovation Institute and chief scientist at Agibot, argues that robotics will not get a "GPT moment" by simply following the LLM development path. Luo says many "embodied foundation models" are effectively mid-training or fine-tuned models rather than true pretrained systems, because high-quality, multi-scenario robot interaction data -- including corrective interventions and long-tail failure cases -- remains scarce. He describes a shift in China's embodied-intelligence sector over the past six months, away from a fixation on robot degrees of freedom and toward data, models, infrastructure, and their interaction in deployment. The piece notes Luo's background at UC Berkeley (under Sergey Levine), Google X, and DeepMind before he returned to China roughly a year ago.
Technical context
Luo's central claim is that the stable statistical relationship between pretraining loss and capability that underpins LLM scaling does not reliably hold for robotics: a declining loss curve mainly shows a model fitting existing data better, not that it can transfer to new physical environments or recover from failure. He points to three technical efforts at Agibot addressing this -- SOP (scalable online post-training infrastructure), LWD (learning while deploying, for continuous post-deployment model updates), and the Tau-0-WM world model, which predicts the physical consequences of candidate actions rather than just generating video. Agibot's concurrent open-sourcing of the AGIBOT WORLD 2026 dataset -- which independently verified reporting puts at just over 1 million real-world trajectories across 217 tasks -- is a concrete instance of the data-infrastructure buildout Luo describes, though the dataset release and the interview are separate items from the same company.
For practitioners
Teams building or evaluating embodied AI systems should treat benchmark and loss-curve improvements with caution absent real-world deployment validation, and should weight investment toward data pipelines that capture failures and corrections rather than only successful demonstrations. Luo's framing suggests procurement and research-prioritization decisions should ask whether a given approach can close a genuine deployment feedback loop, not just whether it scores well on lab benchmarks.
What to watch
Whether other embodied-AI labs publish comparable system-level data-infrastructure work (data standards, shared benchmarks for multi-body interaction data, or cross-robot transfer tooling), and whether Agibot's SOP, LWD, and Tau-0-WM efforts produce measurable deployment gains in the commercial settings Luo cites, such as convenience stores and supermarkets.
Key Points
- 1Agibot chief scientist Luo Jianlan argues robotics lacks a stable loss-to-capability relationship like LLMs, so benchmark gains alone do not predict real-world deployment success.
- 2The binding constraint on embodied AI is scarce multi-scenario robot interaction data, especially failures and corrections, not model or parameter scale.
- 3Practitioners should prioritize data pipelines and deployment feedback loops that capture corrective and long-tail interactions over lab-benchmark optimization alone.
Scoring Rationale
Substantive, well-verified expert commentary from a credentialed practitioner (Luo Jianlan: UC Berkeley under Sergey Levine, ex-Google X, ex-DeepMind, Agibot chief scientist) on a genuine technical bottleneck in embodied AI. Held at 6.5: the core claims check out against independent verification of Agibot's dataset scale and Luo's background, but this remains primarily a single-company interview rather than an independently corroborated industry-wide finding.
Sources
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
