Research

GhostPoint: Self-Supervised Representation Learning by Hallucinating Occluded LiDAR Structure

New research on arXiv explores advanced approaches in robotics and embodied AI, presenting novel methods for robot learning and autonomous manipulation.

Zero-Shot Skeleton-Based Action Anticipation

New research on arXiv explores advanced approaches in robotics and embodied AI, presenting novel methods for robot learning and autonomous manipulation.

Rollplex: Cross-Phase GPU Spatial Sharing for Vision Language Model Post-Training

New research on arXiv explores advanced approaches in robotics and embodied AI, presenting novel methods for robot learning and autonomous manipulation.

SSP: An Event-Matched Syn2Sim2Phy Cross-Domain Evaluation Framework for Autonomous Driving VLA Models

New research on arXiv explores advanced approaches in robotics and embodied AI, presenting novel methods for robot learning and autonomous manipulation.

SparseVoxelDet: Fully Sparse Voxel Networks for Efficient Event-Based Drone Detection

New research on arXiv explores advanced approaches in robotics and embodied AI, presenting novel methods for robot learning and autonomous manipulation.

Beyond Myopic World Models: Long-Horizon End-to-End Training for Direct Future Prediction

New research on arXiv explores advanced approaches in robotics and embodied AI, presenting novel methods for robot learning and autonomous manipulation.

SynthRender and I-AsSET: Open-Source Framework and Dataset for Bidirectional Sim-Real Transfer in Industrial Object Perception

New research on arXiv explores advanced approaches in robotics and embodied AI, presenting novel methods for robot learning and autonomous manipulation.

{\Omega}-QVLA: Robust Quantization for Vision-Language-Action Models via Composite Rotation and Per-step Scaling

New research on arXiv explores advanced approaches in robotics and embodied AI, presenting novel methods for robot learning and autonomous manipulation.

FORGE-plus: Frozen LLM Sets Force Budgets for Contact-Rich Assembly, Passes 256/256 Trials

A two-layer framework lets a frozen text-only LLM assign per-object force ceilings and pick recovery maneuvers from textual force signatures, passing all 256 evaluation episodes on fragile bottle placement and 0.4 mm-clearance gear insertion without breakage.

SoftNav: 3D Scene Tokens Turn a Frozen VLM Into a Transferable Navigator

SoftNav, accepted to IROS 2026, injects entity-level 3D scene tokens directly into a frozen VLM's hidden space — hitting 74.2% success on HM3D-OVON with only ~17M trainable parameters and ~1,200 samples, then walking onto a real Unitree Go2 zero-shot.

KineBench: IDM-Free Benchmark Puts Embodied World Models Through a Kinematic Reality Check

KineBench, accepted to ECCV 2026, evaluates embodied world models without brittle inverse dynamics models: cascaded vision models extract 6D end-effector poses from generated videos and replay them in a physics simulator across 20 ManiSkill3 tasks.

Masked Visual Actions: One Interface Turns Video Models Into Unified Robot World Models

A Stanford-led team introduces Masked Visual Actions, a pixel-space control interface that lets video generation models serve as both forward and inverse dynamics models for robots — finetuned with just 15 hours of masked video examples.