At a symposium in Shanghai on September 15, experts from Chinese robotics and AI companies debated the path to general-purpose embodied intelligence. The consensus: breakthroughs will come not from a single model or technology, but from building complete, scalable systems that combine perception, data, and control.
The event, organized by Gasgoo, brought together executives from Jijia Technology, Fulai New Materials, Guangxiang Lab, Mojia Robot, and Ruierman Intelligent Technology. They discussed how robots can move from understanding the world to acting in it, focusing on model architecture, multi-modal perception, and data strategies.
Model architecture: layered fusion over paradigm wars
A central debate was whether VLA (vision-language-action) models, world models, or physics-native models should serve as the robot's brain. Liao Yongxing, AI algorithm director at Fulai New Materials, argued that the future lies in layered systems rather than a single model type. A primary decision-making "brain" handles high-level semantic understanding and spatial cognition, while a separate "cerebellum" translates motion trajectories into physical interactions.
Liao noted that language models can reason in closed symbolic spaces, but robots must deal with contact, friction, inertia, and real-time control. A single end-to-end model, however impressive in demos, struggles to balance high-level semantics with low-level motor control.
Lyu Yao, chief scientist at Guangxiang Lab, cautioned against paradigm labels altogether. "The paradigm itself shouldn't be the point of contention," he said. The goal is to develop solutions that meet real-world needs, whether the resulting model resembles a VLA or a world model is secondary.
Zhuo Weifeng, head of multi-sensor fusion at Mojia Robot, suggested the brain might not be one unified model but a system of modules for task planning, trajectory generation, and motion control. Ji Haifeng, solutions director at Ruierman, added that VLA models lack physical constraints and next-state prediction, whereas world and physics models have inherent advantages. "Several brains eventually fuse to form the robot's superbrain," he said.
Multi-modal perception and the data bottleneck
As models move toward layered fusion, the next hurdle is integrating touch and force feedback into control loops. Vision alone cannot handle precision tasks like assembly, where contact forces matter. Liao pointed out that touch is often treated as a binary on/off signal, but real feedback involves elastic changes, directional shifts, and object characteristics.
Lyu compared embodied intelligence to human infant learning: through exploration, touching, and grasping, the model collects physical signals. He proposed combining explicit states—like joint angles, velocities, and contact forces—with implicit learning from images, embedding physical dynamics into the training process to reduce data dependence.
Ji illustrated the importance of force sensing in precision assembly. "In a 5mm threaded hole, even a 0.1mm misalignment can prevent insertion," he said. Impedance control requires force and tactile perception to complete such tasks.
Data remains the critical bottleneck. Real-robot data is reliable but costly; simulation data is cheap but must reflect physical laws; human data captures real skills at low cost. Liao believes human data will become more important as simulation matures, with real-robot data used for verification. Lyu warned that simulation should not be seen as a cheap substitute for real data, but as a generator of varied actions under identical conditions, helping models learn underlying physical transitions.
The experts agreed that no single factor—model, data, perception, or hardware—is the sole constraint. "It's a combination," Liao said. "The real challenge is building a closed loop from scenario to model and back to scenario."
