Manipulation Metrics, VLA Distillation, and the Real-Time Control Frontier

Distilling VLA Models to Real-Time Control Speeds

Big VLA models are accurate but slow. This work distills them into fast, lightweight policies that run at 30Hz on standard hardware. The implication is massive: you can finally get foundation-model reasoning in a real-time control loop without a GPU cluster.

MANIPBENCH: A New Standard for Manipulation Evaluation

We're drowning in robot papers with inconsistent metrics. MANIPBENCH proposes a unified benchmark for manipulation, covering perception, planning, and execution across diverse tasks. This is the kind of infrastructure we need to separate real progress from hype.

Foundation Models for Precise, Contact-Rich Assembly

Most manipulation demos are pick-and-place. This paper tackles tight-tolerance assembly, using a foundation model for both high-level planning and low-level force control. It's a direct move toward automating the complex, dexterous tasks that matter in manufacturing.

Tactile Sensing for In-Hand Manipulation Without Vision

Recovering from slips and fumbles requires fine-grained touch. This research shows a policy learned purely from tactile sensors can perform complex in-hand reorientation. It's a clever workaround for scenarios where cameras are occluded or unreliable.

Real-Time Whole-Body Control with Simple State Machines

Getting a humanoid to walk and manipulate simultaneously is a control nightmare. This paper uses a simple state machine to arbitrate between tasks, achieving stable real-time performance. It's a pragmatic, debuggable approach that avoids overcomplicated end-to-end stacks.

Scaling Robot Learning with Cheap, Crowdsourced Demonstrations

High-quality robot data is expensive. This work shows you can train robust policies from noisy, crowdsourced demos collected on cheap hardware. The leverage play is clear: data collection at scale becomes a people problem, not a hardware problem.

3D Diffusion Policy for Generalizable Manipulation

Generalizing to new objects and scenes is the core challenge. This 3D diffusion policy operates directly on point clouds and shows strong zero-shot transfer. It's another step toward manipulation policies that don't break the moment the lab lighting changes.

Closing the Sim-to-Real Gap for Legged Locomotion

Sim-to-real transfer for legs is still brittle. This method uses a novel adaptive mechanism during deployment to correct for simulation errors in real time. It's a tactical fix that makes simulated training dramatically more useful for building real-world quadrupeds.