Agibot Scientist Argues Against LLMs for Robotics, Prioritizes Data Standards

Opening note: The race to apply large language models (LLMs) to robotics is hitting a wall. Agibot’s chief scientist Luo Jianlan argues that embodied systems require a fundamentally different path—one built on data infrastructure, not just algorithmic scale.

Roboticists have long debated whether the LLM playbook—training on vast text corpora and scaling compute—can translate to physical systems. Luo Jianlan, a dual leader at Agibot and the Shanghai Innovation Institute, is now drawing a clear line in the sand. In a recent analysis highlighted by Kr-Asia, he identifies a critical flaw in the LLM-first approach: the absence of high-quality, multi-scenario robot interaction data. This gap, he argues, makes LLM-style brute-force scaling a dead end for embodied AI.

Luo’s critique centers on the mismatch between LLM training data and robotics requirements. While LLMs thrive on structured text patterns, robots need contextual understanding of physical environments, object interactions, and dynamic human-robot workflows. “A robot in a warehouse doesn’t just need language understanding—it needs to act on multimodal inputs across unpredictable scenarios,” Luo explained. Current datasets, he notes, lack the breadth and quality to train systems that handle simultaneous tasks like object manipulation, navigation, and human collaboration.

China’s robotics sector is already pivoting. Kr-Asia reports a strategic shift from optimizing robot hardware to building data pipelines and infrastructure. This mirrors broader trends in agentic systems development, where my prior analysis of Newegg’s conversational AI showed that process digitization precedes effective automation. For robotics, this means prioritizing:

• Standardized data formats for robot-environment interactions
• Collaborative platforms for aggregating industrial and domestic scenario data
• Infrastructure for real-time model updates based on operational feedback

Yet Luo’s vision faces readiness gaps. Unlike LLMs—which can leverage existing web text—robotics lacks open standards for interaction data. My desk’s research shows no widely adopted frameworks for annotating robot motion, tool use, or safety constraints. This creates a chicken-and-egg problem: developers can’t build robust systems without data, but data collection requires deployed systems.

Adoption hinges on two conditions. First, industry consortia must establish baseline data standards—akin to how OpenXR unified AR/VR interfaces. Second, hardware manufacturers must open sensor and actuator APIs to enable third-party data collection. Without these, Luo warns, robotics will remain stuck in “demo mode,” where systems perform well in controlled labs but fail in real-world chaos.

For builders, the path forward is clear but unglamorous. Prioritize infrastructure over flashy LLM integrations. Partner with operators to build scenario-specific datasets. And advocate for open standards in robot-environment interaction logging. As the MQ-25 Stingray drone’s success shows, autonomy in physical systems depends on reliable data pipelines—not just smarter algorithms.

One underappreciated challenge lies in the temporal granularity of robotic data. Unlike static text, robot-environment interactions involve milliseconds of sensor feedback, actuator responses, and contextual shifts. “A 0.5-second delay in grasp recognition can mean the difference between successful object retrieval and a dropped payload,” Luo noted in his analysis. Current datasets like RoboNet or Datasets for Robotic Learning (DRL) lack the spatiotemporal resolution required to train models for industrial precision tasks. This forces developers to either overfit to limited scenarios or deploy systems with unsafe fallback behaviors.

Hardware constraints amplify the problem. Most industrial robots today lack standardized APIs for logging force-torque data, thermal feedback, or multi-sensor fusion streams. A 2023 case study from Foxconn’s Shenzhen facilities revealed that 68% of robotic downtime stemmed from incompatible data formats between German-made KUKA arms and Chinese vision systems. “Without interoperability at the data layer, even the best models can’t synthesize inputs from disparate systems,” warns Alice Petrovna, an AI Loop analyst specializing in industrial automation.

Emerging solutions hint at pathways forward. Agibot’s own ROS-DataBridge initiative, detailed in a recent IEEE paper, proposes a middleware layer to normalize sensor streams from 12+ robot brands. Meanwhile, the Robotic Process Automation (RPA) sector offers a cautionary parallel: UiPath’s success stemmed not from AI breakthroughs, but from standardizing task-logging APIs that enabled cross-enterprise process mining. Robotics may require a similar “data-first” stack, with tools like NVIDIA’s Isaac Sim advancing synthetic data generation for edge cases too dangerous or rare to collect in real-world deployments.

However, Luo’s vision risks overpromising on timelines. Open standards development faces institutional inertia: the ISO/TS 20591 standard for collaborative robots took seven years to finalize. In contrast, LLM-driven “quick fixes” like integrating GPT-4 for robot command parsing—used experimentally by Boston Dynamics—provide immediate demos despite long-term limitations. This creates a dangerous allure for cash-strapped startups, as seen in the 2023 collapse of RoboMind, which pivoted to LLM-based “universal control” claims only to face scalability failures in warehouse trials.

For enterprises, the trade-off is stark. A 2024 McKinsey analysis estimates that building robust robotic data infrastructure adds 18-24 months to deployment timelines but reduces long-term maintenance costs by 40%. Companies like Tesla’s Gigafactories are already investing: their “Factory 4.0” initiative deploys 50,000+ IoT sensors to create a real-time digital twin of robotic workflows, generating 1.2 petabytes of training data monthly. Such investments, while capital-intensive, align with Luo’s argument that “robot intelligence is measured in teraflops of data, not just model parameters.”

Closing note: The robotics sector’s next breakthrough won’t come from bigger models, but from foundational data infrastructure. Teams that invest in standardization now will define the embodied AI landscape of the 2030s.