Guanghui Ren

Robots can learn from fleet-wide failures and successes to autonomously improve their policies, achieving near-perfect reliability without any new human corrections.

Even top-performing robotic value models struggle with nuanced instruction grounding and history-dependent progress tracking, exhibiting inconsistent value judgments across valid execution orders.

Robotic world-action models do not need pretrained internet video backbones to scale: training native planning and dynamics modules from scratch across 30,000 hours unlocks strong zero-shot manipulation and drives a 17.7-point gain on an embodiment comprising under 2% of the data.

MEGA-CL achieves over 75% accuracy in ADMET predictions, significantly outperforming existing models and showcasing its potential for accelerating drug discovery.

Scaling up robot data and closing the loop with state decoding and automated reward scoring allows a 2B parameter video world simulator to outperform larger, dedicated robotic world models in real-world policy transfer.

Decomposing robotic manipulation into coarse and fine-grained actions isn't just conceptually cleaner—it actually unlocks a sweet spot where learning difficulty is balanced, boosting performance.