Yu-Rou Tuan

Part-aware 3D generation can produce parts that look complete but are not physically realizable: neighbouring parts interpenetrate, they have no valid contact surfaces to stay connected, and the assembly falls over in simulation. SNAP3D resolves the inter-part penetration, recovers a contact graph between neighbouring parts, and introduces parameterized connectors at their contact surfaces, then refines each connector’s placement, orientation and dimensions against feedback from physical simulation while preserving the generated geometry. Assemblies are released under gravity exactly as generated, and the exported parts can be printed on a desktop FDM printer and joined by hand without glue or fasteners.

An end-to-end generative framework that synthesizes realistic 4D occupancy and LiDAR data for safety-critical, long-tail driving scenarios. We introduce an entity-centric BEV-to-occupancy method using grounded latents and semantic-Gaussian splatting to preserve rare-class agents during scene edits, and a unified editable bird’s-eye-view interface that integrates real crash records, LLM text specifications, and manual scene edits for controllable data synthesis.

StaDy4D is a large-scale paired static-dynamic 4D benchmark with 9K sequences and 1.9M frames, enabling complete evaluation of dynamic-scene reconstruction in autonomous driving. We propose SIGMA, a test-time-adapted static-dynamic reconstruction method that reduces depth error by 41% over the strongest feed-forward baseline and achieves best-in-class point-cloud reconstruction, with zero-shot transfer to real-world benchmarks.

An OS-level agent-safety benchmark with 82 computer-use tasks across three corrigibility scenarios, evaluating 12 frontier models on OSWorld-Verified. We identify widespread misalignment in benign settings, including 100% override rates and safety failures in spawned subagents, which accessed restricted credentials in 50% of trials.

We use text and visual prompts to enhance pedestrian trajectory prediction, optimizing 2D position retrieval from embeddings for accurate and faster inference. Cross-modal learning improves ADE and FDE performance by over 35%.