Heng Zhang

WorldContact is presented, a contact-centric world model for deformable-object manipulation, constructed from a limited set of high-quality trajectories to generate additional training data efficiently and fine-tune an existing vision-language-action policy and deploy it directly on a real robot.

DeCAL achieves a remarkable 71% success rate in dexterous manipulation tasks by seamlessly integrating tactile sensing with advanced vision-language-action modeling.

Human trajectory logs are no longer the performance ceiling for autonomous driving: closed-loop reinforcement learning paired with distilled foundation models outperforms human demonstration baselines across major open and closed-loop benchmarks.

WDANet outperforms traditional forecasting models by accurately predicting gust peaks during typhoons, crucial for timely disaster response.

Achieving up to 49.4% faster computations for neutron diffusion problems without sacrificing accuracy could revolutionize reactor core analysis.

KENDO achieves up to 5x faster Bayesian optimization and 27x faster active learning while enhancing predictive performance.

Achieving up to 95% success in robotic manipulation tasks, this framework redefines the boundaries of sample efficiency and policy performance in real-world online reinforcement learning.

Entity type information can dramatically boost SciNER performance, enabling LLMs to match fully supervised models without extensive human input.

Automating real-to-sim conversion with vision-language agents could revolutionize how we simulate robotic interactions, making it faster and cheaper than ever before.

BoxTwin enables robots to accurately predict and adapt to the complex dynamics of elastoplastic articulated objects, revolutionizing manipulation in unstructured settings.

NeuralActuator not only predicts actuator dynamics but also enhances sensorless force perception, significantly reducing sim-to-real errors in low-cost robotics.