🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

one-two-vla.github.io Onetwovla Do reasoning when necessary, do action output elsewise

arxiv.org/abs/2505.12679 RL-based ball dribbling for humanoid

zhengyiluo.com/PDC-Site/ Emergent Active Perception and Dexterity of Simulated Humanoids from Visual Reinforcement Learning RL + pulseX prior for visual whole body grasping

arxiv.org/abs/2505.14580 Traversability-aware path planning in dynamic environments Navigation in dynamic polygon cells rather than grids, kinda like graph of convex sets

arxiv.org/abs/2505.17627… H2-COMPACT: Human-Humanoid Co-Manipulation via Adaptive Contact Trajectory Policies

arxiv.org/abs/2505.20573 RLFT LLM to satisfy safety constraints of robotic planning

arxiv.org/abs/2505.19086… MaskedManipulator: Versatile Whole-Body Control for Loco-Manipulation masked mimic + loco mani references, with special treatments of contacts

GET: Goal-directed Exploration and Targeting for Large-Scale Unknown Environments arxiv.org/abs/2505.20828 classic mapping + LLM decision + further traj refinement for exploration.

physicalintelligence.company/research/knowl… VLAs that Train Fast, Run Fast, and Generalize Better do not let the action training pollute the VLA representation.

pku-epic.github.io/TrackVLA-web/ TrackVLA: Embodied Visual Tracking in the Wild simulation data for action model training, vqa data for LM training.

irislab.tech/TwinTrack-webp… TwinTrack: Bridging Vision and Contact Physics for Real-Time Tracking of Unknown Dynamic Objects Combines vision and contact info for pose tracking. First identify geometry and physics, then reuse the physics for prediction.

arxiv.org/abs/2506.01844 smol VLA: smol model, smol training, smol data annotated by smol VLM. But outperforms the big ones.

arxiv.org/abs/2505.24853 DexMachina: Functional Retargeting for Bimanual Dexterous Manipulation Learning to track articulated object states from human demo using RL and curriculum

lecar-lab.github.io/SoFTA/ Hold My Beer: Learning Gentle Humanoid Locomotion and End-Effector Stabilization Control 1. Fast upper body agent + 50hz lower body agent makes a sweet spot for stable loco mani 2. Commands can be used to activate such stabilization

It might be a bit late but here's our work in L4DC 2025 on bridging agility, safety, and adaptivity. Compared to prior work (ABS) that safeguards agile locomotion, this one further explicitly estimates the change in dynamics and does adaptive shielding. adaptive-safe-locomotion.github.io

There is another interesting parallel work from Stanford: arxiv.org/abs/2412.09989 also learns disturbances, and uses learned RA values for shielding. The diff is they use reduced-order model to generate GT reachability values for supervision, while we bootstrap with full model.