Hao Dong, an associate professor at Peking University and chief scientist at Shangwei Qiyuan, argues that the dominant training recipes of imitation learning and reinforcement learning each have hard limits, and the industry needs a different frame.
He splits today’s embodied training into two stages. Pretraining leans on imitation learning, which cold-starts quickly from standardised human demonstrations but trains only on correct trajectories, so the robot never sees failure and cannot self-correct. Post-training leans on reinforcement learning, and the classic Dagger framework, proven earlier in autonomous driving, shows failure samples must be added back so the model learns to recover.
Dong’s lab has already demonstrated a fully autonomous laundry routine. The robot plans its path, opens and closes the washing machine, and on a failed grasp retries like a human rather than stalling, with no human in the loop.
His core idea is a horizontal, two-dimensional scaling law. Beyond the data-volume axis, he adds a task-count axis, so as the dataset grows the initial success rate on new tasks rises and the samples needed per task fall. The target curve is an efficient red line where robot capability expands fast while marginal data cost drops.
Generative data augmentation does the heavy lifting. Using a world model and generative AI, a single real trajectory can yield fifty differentiated, high-fidelity equivalent samples with varied object placement and positions, easing the chronic scarcity and cost of real collection.
Simulation complements real data for non-standard home appliances, and wearable first-person cameras turn human operation videos into cheap robot trajectories. Dong’s conclusion is that only by completing this path do general and home-service robots gain a base for scaled commercialisation.
Editor’s note: This is an adapted translation of the original LeiPhone report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://www.leiphone.com/category/ai/ABnmB3o4JHMsiCmW.html.