Pith. sign in

REVIEW 3 minor 2 cited by

EXPO-FT: Sample-Efficient Reinforcement Learning Finetuning for Vision-Language-Action Models

T0 review · 0 major / 3 minor · reviewed 2026-06-29 · grok-4.3

Pith's one-line read EXPO-FT finetunes pretrained vision-language-action models with reinforcement learning to reach perfect task success using 19.1 minutes of robot data on average.

desk verdict EXPO-FT gets perfect success rates on several precision manipulation tasks with roughly 19 minutes of online data by adding targeted exploration to VLA RL finetuning. read the letter →

arxiv 2605.25477 v2 pith:XNIRCRJK submitted 2026-05-25 cs.RO cs.AI

classification cs.ROcs.AI
keywords reinforcementlearningvision-language-actionmodelsrobotmanipulationsampleefficiencyfinetuningpretrainedpoliciestasks
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Pretrained vision-language-action models generalize across manipulation tasks yet fall short on the reliability needed for deployment. EXPO-FT applies reinforcement learning to fine-tune these models in a stable and sample-efficient way. The approach is tested on tasks that combine high precision, dynamic movements, and varied starting positions, such as routing string lights, striking a pool ball, and inserting a flower into a bottle. It reports perfect success rates across the evaluated suite while using far less online data than training from scratch or prior finetuning methods.

What carries the argument

EXPO-FT, the system that performs stable reinforcement learning fine-tuning on pretrained vision-language-action policies

What would settle it

Recording fewer than 30 successes in 30 trials or requiring substantially more than 19.1 minutes of online data on average for the pool ball striking or flower insertion tasks.

Watch

Extended reading notes

Core claim

EXPO-FT is a system for stable, sample-efficient RL finetuning of pretrained VLA policies that solves a suite of challenging manipulation tasks, including routing string lights and inserting the plug to light it up, striking a pool ball into a pocket, and inserting a flower into a wine bottle, each requiring combinations of high precision, dynamic actions, and robustness to varied initial states. Our system achieves perfect task performance (30/30 successes) across all evaluated tasks within an average of 19.1 minutes of online robot data, outperforming both prior RL-from-scratch and VLA finetuning approaches.

Load-bearing premise

That the EXPO-FT system can deliver the claimed stability and sample efficiency on the described suite of high-precision, dynamic manipulation tasks when applied to pretrained VLA policies.

Editorial extensions

If this is right

  • Pretrained VLA policies reach perfect success rates on high-precision tasks after limited online interaction.
  • The method uses less data than RL trained from scratch while improving on prior VLA finetuning results.
  • Tasks that combine dynamic actions with robustness to initial state changes become reliably solvable.
  • An open-source release supports wider testing of RL finetuning for VLA models in robotics.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same finetuning pattern could be examined on tasks outside tabletop manipulation, such as mobile navigation or multi-arm coordination.
  • If efficiency scales, the approach might reduce the total pretraining data needed by shifting more adaptation burden to short RL stages.
  • Testing on hardware with greater sensor noise or longer task horizons would reveal whether the reported data requirements remain stable.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

0 major / 3 minor

Summary. The manuscript introduces EXPO-FT, a system that augments RL finetuning of pretrained Vision-Language-Action (VLA) policies with an exploration objective, reward shaping, and VLA adaptation procedure. It reports solving a suite of high-precision, dynamic manipulation tasks (routing string lights, striking pool balls, inserting flowers into bottles) to 30/30 success using an average of 19.1 minutes of online robot data per task, outperforming both RL-from-scratch and prior VLA finetuning baselines.

Significance. If the reported outcomes hold under the stated data budgets and task conditions, the work provides a concrete route to reliable real-world deployment of VLAs by addressing stability and sample-efficiency gaps. The open-source codebase release is a clear strength that supports reproducibility and adoption.

minor comments (3)
  1. The experimental section should explicitly state the number of independent random seeds or rollouts used to compute the 30/30 success rates and any associated variance, to strengthen the stability claim.
  2. Figure captions and baseline descriptions would benefit from additional detail on hyperparameter matching across methods to ensure fair comparison.
  3. A short discussion of failure modes or edge cases observed during the 19.1-minute finetuning runs would improve clarity on the method's robustness limits.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their positive assessment of EXPO-FT and for recommending minor revision. We appreciate the recognition that the reported outcomes, if they hold, provide a concrete route to reliable real-world VLA deployment, as well as the value placed on the open-source codebase.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity identified

full rationale

The paper is an empirical robotics contribution describing an RL finetuning system (EXPO-FT) and reporting experimental success rates on manipulation tasks. No derivation chain, first-principles predictions, fitted parameters renamed as outputs, or self-citation load-bearing steps appear in the provided abstract or described method/experimental sections. Claims rest on reported robot trials rather than any definitional or self-referential reduction.

Assumptions & free parameters 0 free parameters · 0 assumptions · 0 invented entities

Only the abstract is available; no technical details on parameters, axioms, or entities are provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EXPO-FT: Sample-Efficient Reinforcement Learning Finetuning for Vision-Language-Action Models." pith.science (2026). https://pith.science/paper/XNIRCRJK

@misc{pith2026260525477,
  author       = {Pith},
  title        = {Pith review of: EXPO-FT: Sample-Efficient Reinforcement Learning Finetuning for Vision-Language-Action Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XNIRCRJK}},
  note         = {Machine review of arXiv:2605.25477}
}
read the original abstract

The ability to efficiently and reliably learn new tasks has been a foundational challenge in robotics. Vision-Language-Action (VLA) models have demonstrated strong generalization across diverse manipulation tasks, yet pretrained policies consistently fall short of the reliability required for real-world deployment. Reinforcement learning (RL) fine-tuning offers a promising path to bridge this gap, but existing approaches either train from scratch without fully leveraging pretrained priors, or fine-tune VLAs without achieving the sample efficiency and success rates that practical deployment demands. We present EXPO-FT, a system for stable, sample-efficient RL finetuning of pretrained VLA policies that closes this gap. Our system solves a suite of challenging manipulation tasks, including routing string lights and inserting the plug to light it up, striking a pool ball into a pocket, and inserting a flower into a wine bottle, each requiring combinations of high precision, dynamic actions, and robustness to varied initial states. Our system achieves perfect task performance (30/30 successes) across all evaluated tasks within an average of 19.1 minutes of online robot data, outperforming both prior RL-from-scratch and VLA finetuning approaches. We release an open-source codebase with the aim of facilitating broader adoption of RL finetuning of VLA models in robotics.

Figures

Figures reproduced from arXiv: 2605.25477 by the authors.

Figure 1
Figure 1. Average training success rates of EXPO￾FT compared to prior methods. EXPO-FT achieves a reliable performance with high sample efficiency where prior methods often do not converge reliably. We empirically find that our system achieves dexterous and precise manipulation capabilities across a diverse set of challenging tasks, includ￾ing routing string lights and inserting the power connector to illuminate them, strikin… view at source ↗
Figure 2
Figure 2. Left: Overview of EXPO-FT. EXPO-FT features a server that handles VLA training and inference and a learner process that steps in the environment to enable VLA finetuning with RL. Right: Architecture of EXPO-FT. EXPO-FT finetunes the VLA model with EXPO for sample-efficient training. cases. We start by describing the problem statement (Section 4.1), then describe the approach used for finetuning (Section 4.2), and la… view at source ↗
Figure 3
Figure 3. Eight real-world manipulation tasks in our evaluation suite. Flower Insert (tight insertion tolerances), String Light Routing - RouteI/II, Insert (long-horizon precise alignment), Egg Flip (dynamic contact￾rich tool use), Candy Scoop (stable control in visually messy scenes), Pool Shot (precise speed control) and Cube Pick (large scene randomization). The tasks span dexterous, precise, deformable, and dynamic manipu… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Training success and intervention rates across all tasks. Top row: Egg Flip, Flower Insert, Pool Shot, Cube Pick. Bottom row: String Light Routing - Route I, String Light Routing - Route II, String Light Routing - Insert, Candy Scoop [PITH_FULL_IMAGE:figures/full_fig_…
Figure 5
Figure 5. Figure 5: Episode Time across all tasks. Top row: Egg Flip, Flower Insert, Pool Shot, Cube Pick. Bottom row: String Light Routing - Route I, String Light Routing - Route II, String Light Routing - Insert, Candy Scoop. B Detailed Task Setting B.1 Task Setting Description Here, we…
Figure 6
Figure 6. Figure 6: Task strips demonstrating successful completion of each task. Candy Scoop. We pre-collect 20 demonstrations for this task. The reward classification for this task is split into two parts, both of which must succeed for the episode to be counted as successful. In the fi…
Figure 7
Figure 7. Figure 7: Visualization of randomized initial state spaces for all tasks. The orange regions indicate the randomized initialization areas used during training. C Detailed Training Setting C.1 Model Structure/Training Detailed We instantiate EXPO-FT with π0.5 [1] as the base poli…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Do You Really Need to Pretrain Q-Functions for Online RL Fine-Tuning?

    cs.LG 2026-07 conditional novelty 6.0 of 10

    On six robot-manipulation tasks, offline Q-pretraining does not accelerate online RL fine-tuning from a pretrained policy, while seeding the replay buffer with rollouts from an ensemble of policies (IPE) improves fina...

  2. VistaVLA: Geometry- and Semantic-Aware 3D Gaussian-Grounded VLA for Robotic Manipulation

    cs.RO 2026-07 conditional novelty 5.5 of 10

    VistaVLA lifts multi-view vision-language features into 3D Gaussians, compresses them 99% via Merge-then-Query, and improves real-robot manipulation success by ~23% over baselines.

Pith tools

Reviewed June 29, 2026 · model on record in the stance chip above.