Cost Curves, Composition, and the VLA Efficiency Wars

Efficient VLA via Speculative Decoding and Multi-Scale Attention

This paper attacks the inference latency problem head-on, claiming massive speedups using speculative decoding. If the performance holds up in messy real-world scenes, this is a direct unlock for deploying VLAs on edge hardware without melting the battery.

Compositional Robot Planning with Vision-Language Models

Stop trying to force-feed everything into one massive end-to-end model. This work focuses on teaching robots to decompose tasks using VLMs, which is a much more practical path to getting high reliability on long-horizon jobs.

Distilling VLA Policies to 30Hz for Real-Time Control

Getting Vision-Language-Action models up to a 30Hz control loop is the holy grail for dynamic tasks. This distillation approach looks promising for builders who need VLA reasoning without the slow loop times that usually come with it.

REBEL: A Benchmark for Real-Time Robot Learning

We are drowning in new models but starving for reliable metrics. A standardized benchmark for real-time learning is critical for builders to compare apples to apples and stop relying on cherry-picked demo videos.

Humanoid Locomotion via Simplified Dynamics and RL

Sometimes over-complicating the dynamics model just slows down training. This paper suggests stripping things back to a simplified model can actually train more robust locomotion policies that transfer better to hardware.

Few-Shot Sim-to-Real Transfer via Task-Oriented Adaptation

Sim-to-real transfer is usually a pain, but adapting specifically to the task features instead of the whole scene cuts down the domain randomization nightmare. A huge time saver if you're tuning for a specific pick-and-place line.

LEO: Learning to Explore and Optimize for General Manipulation

Exploration strategies for general manipulation are usually random or scripted, which scales poorly. Teaching a robot to actively explore objects to find optimal grasp points sounds like a win for reducing the engineering time on new SKUs.