Investigates why VLA models (e.g., π0.5) fail dramatically under novel viewpoints. Decouples the issue into physical vs. spatial modeling failures, showing that the pretrained model retains strong physical understanding while spatial representation mismatch is the key bottleneck. Proposes two lightweight adaptation methods — FTM (token affine modulation) and FLA (low-rank ViT update) — that recover cross-viewpoint performance by updating only 4K–4.7M parameters, achieving 90.8% success rate on the LIBERO-V benchmark with a 99× parameter efficiency gain over LoRA. One-shot sim-to-real transfer validated on a real Franka arm.

Proposes HumanGenesis, a multi-agent collaborative framework that unifies Real2Sim and Sim2Real in a closed loop for human dynamics modeling. The system integrates a 3DGS+SMPL+Learnable-LBS Reconstructor, a Qwen2.5-VL-driven Critique Agent with multi-round self-reflection for fine-grained reconstruction refinement, and a Video Harmonizer that enhances human-scene consistency and temporal coherence in rendered videos. Achieves state-of-the-art results on HumanVid and NeuMan benchmarks.

Proposes Attention-Conditional Diffusion (ACD), a controllable video generation framework built on CogVideoX. Unlike traditional guidance-level conditioning, ACD directly supervises cross-attention maps inside the diffusion model via a dual-branch (masked/unmasked) shared-parameter fine-tuning scheme, moving conditioning from output-level to attention-level and eliminating common artifacts. Uses sparse 3D-aware object layouts as control signals with a layout ControlNet, supported by an automated annotation pipeline on 20K RealEstate training clips. Outperforms AC3D and other baselines on FID/FVD and camera error metrics.