Robot Brain Race Enters Deep Water: Why Physical World Requires Designing Models from Scratch

In an interview at WAIC, Ant Lingbo's chief scientist Shen Yujun discussed the challenges of embodied intelligence, emphasizing that data standards have not converged because technical routes are still undecided, with no consensus on modalities, precision, or quality. He noted that real data, especially teleoperation data from robots and first-person data from humans, is more critical than simulation data for general-purpose robots, as simulations struggle to replicate human subconscious actions. Lingbo's strategy involves developing multiple models (like LingBot-Vision, LingBot-Depth, LingBot-Video, LingBot-VLA, and LingBot-VA) to solve specific technical problems, such as aligning modalities, predicting dynamics, and ensuring load balancing in MoE architectures, rather than attempting a single end-to-end model prematurely. Shen introduced the concept of 'embodied-native' models, requiring data native to physical world tasks, model capabilities suited for real-time interaction, and the ability to train from scratch, citing difficulties in replicating self-supervised vision models like DINO and implementing causal modeling for one-way attention. On safety, he stressed that 'fence-based' safety (restricting actions) is insufficient, advocating for 'native safety' embedded during pretraining, similar to human instincts. He also predicted that the 'ChatGPT moment' for embodied AI will occur when ordinary people can easily participate in data collection, making robots truly beneficial and integrated into daily life.