VLA Patching, Constrained Safety, and Foundation Locomotion

Language-Grounded Whole-Body Control for Bipedal Loco-Manipulation

Unified verbal grounding + Locomotion controller stack. Worth watching if you're building humanoids that need to navigate cluttered real spaces.

Diffusion-Guided Sampling for Long-Horizon Robot Planning Efficiency

Combines diffusion objectives with policy rollouts for better test-time compute allocation. Useful trick for mobile manipulators with battery constraints.

Vision-Language-Action Pipeline Patching via Retrieval-Augmented Corrections

Drop-in two-stage module that patches VLA brittle patches without full retraining. Saves weeks of GPU time when your manipulation policy hallucinates on novel objects.

Unpaired Bimanual Contact Distillation from Demonstrations

Shows how to distill contact-rich bimanual skills from teleop without paired clean data. Practical if you're scaling up manipulation demos cheaply.

Geometric Attention for Object-Centric Robotic Manipulation

Propagates geometric structure into transformer attention for tighter robot-object correspondence. Good signal for pick-and-place generalization benchmarks.

Constrained Flow Matching for Real-Time Trajectory Safety Guarantees

Combines flow-matching trajectory generation with constrained optimization for safe arm control. Worth stealing if you're deploying robots near human workers.

Foundation Feature Transfer for Zero-Shot Legged Locomotion

Leans on foundation model features to bootstrap legged robot locomotion in unseen terrain. Reduces sim-to-real mass specs from days to hours.

Synthetic Video Augmentation Limits for Manipulation Generalization

Benchmark of how much video augmentation helps without poisoning real-world performance. Useful baseline for anyone over-fitting on synthetic data.