Embodied failure recovery
Collecting auditable trajectories that preserve failures, recovery attempts, and counterfactual branches.
Research
Page 1 - Showing 3 of 3themes
I am interested in building agents that can learn from interactions, failures, and visual dynamics rather than only from final outcomes.
Embodied AIWorld ModelsVideo UnderstandingMixture-of-ExpertsData-Centric Learning Systems
Video world models
Studying how routed video models represent visual changes, dynamics, and physical interactions.
Multi-agent data
Contributing interaction data and quality control for world-model research in multi-agent settings.
Current work
Work in progress, stated plainly.
ICLR 2027 submission · arXiv:2609.37829
DIET: DeletIon-response Expert Trimming for Video Diffusion Transformers
Training-free pruning of 50% of the experts in a 30B video diffusion transformer: the checkpoint footprint drops from 57 GB to 30 GB, enabling single-card deployment on a 48 GB GPU, with the official VBench Total rising from 0.7941 to 0.8115 on a fixed 284-case protocol.
Project pageAAAI submission in preparation
Embodied Failure-Recovery Data Collection
A data-collection and evaluation workflow for recoverable embodied task failures in AI2-THOR, with replayable traces and branch-level evidence.
GitHubOngoing collaboration
Multi-Agent World-Model Data Collection
Contributing multi-agent interaction data and quality control. Model training is planned for a later stage.