DexVLG: Dexterous Vision-Language-Grasp Model at Scale. Object segmentation -> optimized grasp per segment and VLM generated caption -> big dataset to train VLG model

Compliant Residual DAgger: Improving Real-World Contact-Rich Manipulation with Human Corrections. human correction as the expert signal in dagger to finetune the policy, supported by compliant hardware design.

Learning Accurate Whole-body Throwing with High-frequency Residual Policy and Pullback Tube Acceleration. low-freq nominal EE tracking policy + high-freq arm correction policy + Backward Reachable Tube based acceleration commands

Dex1B: Learning with 1B Demonstrations for Dexterous Manipulation. 1. optimization-based dataset initialization.2. CVAE-based policy training with pointnet encoder.3. increase the diversity by repetitively sampling rare cases and finetuning the policy

Versatile Loco-Manipulation through Flexible Interlimb Coordination. 1. a mask per limb for manipulation or locomotion.2. whole body IK to get manipulation and torso goals.3. RL for loco balancing, model based for manipulation tracking

Eye, Robot: Learning to Look to Act. BC arm + RL eye for active perception

From Experts to a Generalist: Toward General.Whole-Body Control for Humanoid Robots. training separate experts for different clusters of motions, (also treat sim2real separately), then distill to one WBC.

kungfu bot. motion filtering + adaptive tracking rewards for kungfu motion tracking

RL from Physical Feedback: Aligning Large Motion Models with Humanoid Control. use physical simulation to verify and finetune text2motion models

CASPER: Inferring Diverse Intents for Assistive Teleoperation with Vision Language Models.

"This unified representation significantly reduces the need for extensive data collection for each new robot morphology, accelerates generalization across embodiments, and ultimately facilitates more scalable and efficient robotic learning.".

Latent Action Diffusion for Cross-Embodiment Manipulation. shared latent space DP with different en/decoders for different morph

Narrate2Nav: Real-Time Visual Navigation with Implicit Language Reasoning in Human-Centric Environments. teacher VLA (with language narration of scene as input) + student distillation to implicitly learn the cues with real-time inference.

Vision in Action: Learning Active Perception from Human Demonstrations. Title tells what is done. The system is specially optimized for such kind of things.

RT @ChongZitaZhang: There is another interesting parallel work from Stanford:. also learns disturbances, and uses l….

RT @ChongZitaZhang: It might be a bit late but here's our work in L4DC 2025 on bridging agility, safety, and adaptivity. Compared to prior….

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

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

smol VLA: smol model, smol training, smol data annotated by smol VLM. But outperforms the big ones.

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.