Ph.D. Student @ Tsinghua University
I work on robot learning and embodied intelligence, with a focus on VLA policies, generative world models, and data-efficient robot manipulation.
I am a fourth-year direct-entry Ph.D. student in Electronic Information at Tsinghua University (SIGS), advised by Prof. Xiu Li in the Intelligent Computing Lab. I study how generative models support physical decision making through spatial grounding, data-efficient diffusion policies, and task-conditioned priors from frozen video models.
One paper, More than A Point: Capturing Uncertainty with Adaptive Affordance Heatmaps for Spatial Grounding in Robotic Tasks, was accepted to CoRL 2026.
One paper, Beyond Video Generation: Exploring Latent Interaction Priors from Frozen Video World Models for Robot Learning, was accepted to the NeurIPS 2026 Workshop: Robot Learning with World Models: Capabilities, Frontiers, and Challenges.
One paper, RASR: Range-Aware Scale Recovery for Metric UAV Navigation, was accepted to the ACM MM 2026 Workshop UAVM.
One paper, Linear Differential Vision Transformer: Learning Visual Contrasts via Pairwise Differentials, was accepted to NeurIPS 2025.
CoRL 2026 · Accepted
Adaptive affordance heatmaps represent continuous spatial uncertainty instead of brittle points or boxes, improving robustness to ambiguous instructions and perceptual noise in real-world manipulation.
arXiv preprint · 2026
A single-step latent velocity field from a frozen video model provides task-conditioned where-and-how cues for low-data manipulation, reaching 90.6% on LIBERO and 65.7% on RoboTwin 2.0 Easy.
arXiv preprint · 2025
A function-centric taxonomy organizes generative decision-making methods as Controllers, Modelers, Optimizers, and Evaluators within the Control-as-Inference framework.
arXiv preprint · 2026
Spatial visual prompts decouple semantic reasoning from geometric grounding by turning language instructions into explicit zero-shot masks for a lightweight imitation policy.
ACM MM 2026 Workshop UAVM · Accepted
RASR recovers metric distance and heading from frozen pairwise geometry using global calibration and range-aware residual correction for GNSS-denied UAV navigation.
NeurIPS 2025 · Accepted
Visual-Contrast Attention uses pooled positive and negative token streams to sharpen discriminative modeling while keeping the attention computation linear and adding no extra FLOPs.