TAU-Bench is a track-centric benchmark for jointly evaluating anomaly instance tracking and fine-grained video anomaly understanding.

ChainVLA is a 1.2B-parameter VLA policy that connects successive action queries through a joint, revisable execution state, carrying both task progress and unfinished motion across long-horizon manipulation.

DynTrace is a training-free framework that tracks continuous dynamic object evidence for coherent 4D reasoning. It combines geometry-informed trajectory visualization with a structured Dynamic Trace Graph.

RoboStream weaves spatio-temporal reasoning with persistent memory into VLM robotic planners, supporting long-horizon object grounding and causal state tracking.

RobotWorld investigates how general-purpose multimodal agents transfer their digital tool-use capabilities to physical task execution through robot interfaces. Its 84 simulation tasks span manipulation, mobile manipulation, locomotion, driving, and aerial control, with explicit interaction budgets and executable success checks. Agents can construct sophisticated perception and control workflows involving segmentation, camera calibration, spatial estimation, and dynamics computation, yet these capabilities do not consistently compose into successful behavior. Execution-trace analysis exposes lost object states, ineffective action correction, late recovery, and mistaken completion judgments, providing concrete targets for more reliable physical-world agents.

Robot foundation policies predict action chunks, but how many actions to execute before replanning depends on the current task phase. ChunkTrust treats the execution horizon as a latent variable inferred from action-expert evidence. Its training-free Action-aware Horizon Selector (AHS) combines intra-chunk spectral stability with inter-chunk motion continuity, while an online Beta posterior with kernel forgetting tracks horizon preferences across replans. An optional Query-based Horizon Adapter (QHA) learns a context-conditioned prior while keeping the base policy frozen. AHS improves success across evaluated RoboTwin2.0 and RoboCasa GR1 Tabletop configurations, and raises the mean normalized process score on four real-world household tasks from 50.4% to 57.5%.

Zip-VGGT introduces object-centric spatiotemporal KV compression for streaming vision, preserving geometric topology and dynamic trajectories while controlling long-term memory use.

Dyn-Bench evaluates whether MLLMs can perceive, track, and reason about evolving 4D scenes using 1K videos, 7K VQA pairs, and 3K dynamic-object grounding pairs.

DynamicVerse is a physical-scale multimodal 4D modeling framework for real-world video, combining metric geometry, motion, masks, and descriptive captions.

TAO reframes granular video anomaly detection as pixel-level tracking of anomalous objects, linking anomaly scoring with segmentation and tracking for precise long-video localization.

This work turns SAM's prompt sensitivity into an advantage for ambiguous segmentation, enabling diverse, plausible, and controllable predictions.

P²SAM models ambiguous medical-image segmentation through a probabilistic prompt space built on SAM, improving accuracy and diversity with limited annotations.