Reproduce “$μ$VLA: On Recurrent Memory for Partially Observable Manipulation in VLA Models” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Egor Cherepanov, Nikita Kachaev, Daniil Zelezetsky, Aydar Bulatov, Artem Pshenitsyn, Yuri Kuratov, Alexey Skrynnik, Aleksandr I. Panov and 1 more

μVLA adds 64 learned memory tokens to an existing vision-language-action transformer, updating its state from each observation and using it with the current view. With truncated backpropagation through time, it raises mean success across five MIKASA-Robo training tasks to 0.84, versus 0.48 for an episodic memoryless reference. Held-out gains are smaller, and novel memory tasks remain difficult.

Key insights

  • Minimal recurrent memory can be effectively added via fine-tuning to a memoryless VLA, yielding substantial performance improvements on partially observable tasks without retraining the full backbone.
  • The optimal Truncated Backpropagation Through Time (TBPTT) length for recurrent training exhibits a U-shaped performance curve, with K=2 providing the best balance for credit assignment across timesteps.
  • While effective for matched memory semantics, the minimal in-backbone recurrence approach shows limitations in generalizing to held-out tasks with conceptually novel memory requirements.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.