Fri 13 Nov
LiveGet it free
ZOTPAPER

News without the noise

← All concepts

vision-language-action model

A vision-language-action (VLA) model lets a robot see, understand language, and act. It processes camera images and text instructions to decide what to do next, often predicting a short sequence of future actions called an action chunk. The model then replans after executing that chunk. Making these models robust means they can handle corrupted sensor data, adapt to different situations, and respond quickly to hazards without extra supervision.

The papers focus on making VLA models adapt their own decision process. Some methods adapt the action chunk length based on prediction confidence or scene risk, improving success under corrupted observations and in hazardous driving. Others use language as a cheap memory or synthetic demonstrations to replace human data, rescuing tasks with zero initial success. Evaluations reveal that current models struggle with spatial reasoning and long-horizon planning, highlighting the need for better adaptability and robustness.

Papers
84
Released code
3
First seen
Aug 2026
Latest
Nov 2026

80 papers in the last two months, against 4 in the two before.

Start here

Three papers to read first, in the order they came out
  1. 1AI driving models go farther when action feedback updates perception

    The benefit of VLM gradient insulation depends on action-token supervision and evaluation, questioning its default use in VLA training.

    Nov 2026
  2. 2Simple retrieval beats most trained robot policies on a popular benchmark

    Nearest-neighbor reuse of demonstration actions reaches 91.7% on LIBERO, showing that much benchmark success can come from matching familiar states rather than learned control.

    Nov 2026
  3. 3Robot memory that runs by clock confuses waiting with working

    Time-sampled memory makes robots mistake idle waiting for task progress, causing early stopping during null events.

    Nov 2026

How fast it is moving

JunJulAugSeptOctNov so far

Often studied with

world model for agentsA predictive environment model that agents use to simulate outcomes, enabling planning, reasoning, and long-horizon task execution.22 togetherflow matchingA generative training method that learns an ordinary differential equation mapping noise to data by matching velocity fields along probability paths.22 togetheragent memoryA repository where an agent stores past interactions and outcomes to inform future decisions.13 together

Who is working on it

NVIDIA 8University of Washington 4National University of Singapore 4KAIST 4University of Chinese Academy of Sciences 4Peking University 3

The papers

Most central to this idea first, not most recent.

Concepts are extracted from each paper and reused across the corpus, so this page grows on its own as the desk reads.