Wenchao Ding

Task execution time can be drastically cut while improving success rates by adapting motion speeds to the complexity of manipulation stages—no annotations required.

StreamKL slashes memory usage from quadratic to constant, enabling efficient long-context attention distillation on a single GPU.

TacForeSight enables robots to anticipate contact changes in real-time, outperforming traditional methods in dynamic manipulation tasks.

Robots can now perform complex, contact-rich tasks with significantly smoother and more continuous motions by learning high-frequency action chunks in a latent space.

Autonomous vehicles can now better navigate occluded environments thanks to a new risk map that understands both traffic flow and collision potential.

Achieve robust robot manipulation across diverse viewpoints without camera calibration by synthesizing novel views with a geometry-aware video diffusion model.

Pocket-sized VLA models can now achieve state-of-the-art robot manipulation performance by pre-training on a curated multimodal dataset and injecting manipulation-relevant representations into the action space.

End-to-end driving models are surprisingly bad at using navigation, but a new framework shows how to inject it for SOTA results.

LLMs can now write better hardware verification code: CoverAssert boosts functional coverage by up to 15% by iteratively guiding LLMs with coverage feedback.

Humanoid robots can now traverse steps up to 20cm and gaps up to 80cm thanks to a contrastive learning approach that forces MoE experts to specialize on different terrain types.

Watch two robots hug and dance: Rhythm enables robust, physically plausible interactions between dual humanoids, bridging the gap from simulation to real-world deployment.

Robots that learn from their mistakes *while* navigating? SERP unlocks this by evolving the action model in-context during replanning, boosting success rates and cutting token costs.