ChunkVLA-AM: Parallel Action Chunking for Vision-Language-Action Robot Control in Additive Manufacturing

Zhugang Liu; Kaichuang Zhang; Jinman Zhang; Pu Sun; Martha Asare; Jose Hernandez; Maxim Ermolinsky; Efren Saenz

cs.ROOct 1, 2026

ChunkVLA-AM: Parallel Action Chunking for Vision-Language-Action Robot Control in Additive Manufacturing

Authors: Zhugang Liu, Kaichuang Zhang, Jinman Zhang, Pu Sun, Martha Asare, Jose Hernandez, Maxim Ermolinsky, Efren Saenz, +2 more

Organizations: Department of Computer Science, The University of Texas Rio Grande Valley · Department of Electrical Engineering, University of South Florida · Department of Computer Science, San Diego State University · Department of Electrical and Computer Engineering, The University of Texas Rio Grande Valley

Abstract

Vision-language-action (VLA) models unify visual perception, language understanding, and action generation, offering new opportunities for automation in additive manufacturing (AM). However, deployment in AM remains challenging because adapting these models to unseen robot embodiments is costly, and performance can degrade under environment changes. In this work, we present a framework for deploying OpenVLA-OFT on a FAIRINO FR3 robot in a fixed AM workcell. A data pipeline converts monocular real-world demonstrations into OpenVLA-compatible TFDS/RLDS datasets to support adaptation to the FR3 embodiment. At runtime, each inference request predicts an eight-step chunk of 7-D actions. The FR3 executes each chunk open loop before capturing a new observation, providing closed-loop feedback between chunks. The system uses a cloud-edge architecture in which the FR3 client streams observations to a remote inference server through a FastAPI interface. In 42 physical A-to-B object-transfer trials, evenly split between red and blue targets, the system succeeded in 39 (92.9%). All three failures occurred during final placement, when insufficient release-height control caused the object to topple. An illumination sweep identified a low-error luminance range of 85-125 on a 0-255 scale, with the lowest mean spatial error at 95.

Figures & tables

Fig. 1: Cloud-edge deployment architecture. The local FR3 client sends instruction-conditioned RGB observations to a remote OpenVLA-OFT service, receives an action chunk, and validates commands before execution.

Fig. 2: Trajectory tracking comparison across adaptation methods. The y-axis uses a logarithmic scale to highlight the significant precision gains. ChunkVLA-AM achieves sub-millimeter lateral accuracy, outperforming both zero-shot models and single-step fine-tuned baselines.

Fig. 3: Trajectory prediction errors across ten luminance-sweep trials. The solid curve and surrounding band show the across-trial mean spatial error and $\pm 1$ standard deviation, respectively. The remaining curves show the axis-wise MAEs. The highlighted range from 85 to 125 indicates a low-error region centered around the minimum error observed at luminance 95.

Panel of Figure 4 — Fig. 4: Representative FR3 workspace captures under six illumination conditions used in the lighting robustness study. Panel (a) is the neutral baseline; the remaining panels span low light, over-exposure, and warm/cold color-temperature variants. Luminance values are mean pixel intensity on a 0–255 scale.

Condition	Luminance	MAE (mm)
	(0–255)	$x$	$y$	$z$	3D $L_{2}$
Baseline (Neutral)	109	0.37	0.97	3.87	4.01
LowLight Extreme	30	0.45	1.28	4.04	4.26
LowLight Mild	65	0.26	1.03	4.03	4.17
OverLight Mild	165	0.37	1.00	3.86	4.00
OverLight Extreme	190	0.59	2.30	4.65	5.22

TABLE I: Influence of environmental illumination and color temperature on chunked trajectory prediction. Brightness is measured in mean luminance on a 0–255 scale. Errors are reported in axis-wise Mean Absolute Error (MAE) alongside the 3D Euclidean error (

L_{2}

Luminance region	Sweep error	Axis MAE (mm)
(0–255)	(mm)	$x$	$y$	$z$
30–84	5.872	0.834	1.209	4.985
85–125	5.499	0.676	1.198	4.639
126–210	5.644	0.720	1.331	4.688

TABLE II: Trial-level summary of the dense luminance sweep. The reported sweep error is the millimeter-valued spatial-error metric provided in the analysis outputs and is kept distinct from the 3D

L_{2}

values in Table I . Region means are first computed within each of the 10 trials.

Explore similar work

When to Switch: Reliable Action-Chunk Extension for Vision-Language-Action Models
Oct 5, 2026Seonghoon Yu, Dongwon Kim, HyungRok Jung +4
KAIST · GIST · POSTECH +1
VOTE: Vision-Language-Action Optimization with Trajectory Ensemble Voting
Jul 7, 2025Juyi Lin, Amir Taherin, Arash Akbari +11Diffusion-Based Vision-Language-Actions Robotic Manipulation
Northeastern University, Boston, USA · EmbodyX,San Mateo,USA
On the Efficiency of LoRA Fine-Tuning for Vision-Language-Action Models in Industrial Robotic Manipulation
Jul 11, 2026Finn Ferchau, Daniel Pommer, Cristian AxenieDiffusion-Based Vision-Language-Actions Flow-Matching Vision-Language-Action
Technische Hochschule Nürnberg Georg Simon Ohm, Nürnberg, Germany · Siemens AG, München, Germany · Fraunhofer Institute for Integrated Circuits (IIS), Erlangen, Germany

Oct 5, 2026, cs.ROJ/K move · Enter open · S save

When to Switch: Reliable Action-Chunk Extension for Vision-Language-Action Models

Seonghoon Yu, Dongwon Kim, HyungRok Jung, Yoonjae Baek+3

KAIST · GIST · POSTECH+1

Open paper AskPDF arXiv

Vision-Language-Action (VLA) models serve as unified policies for robotic manipulation, yet their expensive inference forces robots to pause between policy calls, resulting in stop-and-go execution that interrupts smooth motion and prolongs task completion. Extending the action chunk reduces policy calls and hence these pauses, but predicting farther into the future makes long-chunk execution unreliable. To understand where this unreliability arises, we analyze action errors within long chunks and find that they concentrate around transitions between manipulation subskills, growing sharply with chunk length. This suggests the importance of transition timing, i.e., when to switch subskills within a chunk. Motivated by this observation, we introduce RACE (Reliable Action-Chunk Extension), a framework that predicts the transition timing from an auxiliary one-step denoising pass and conditions action generation on it. By learning and conditioning on transition timing, RACE reduces errors at subskill transitions and enables reliable execution of longer action chunks. Across simulation benchmarks, RACE outperforms fine-tuning at the same chunk length; with 2x longer chunks, it surpasses recent state-of-the-art and efficient VLAs in success rate, and with 4x longer chunks, it remains competitive. On a real robot, RACE uses 4x longer chunks, which reduces the idle time caused by stop-and-go execution by about 5x, while achieving a higher success rate than fine-tuning with the same chunk length. Code and a real-robot demo are available at https://github.com/Seonghoon-Yu/RACE-VLA

Abstract

Explore similar work

When to Switch: Reliable Action-Chunk Extension for Vision-Language-Action Models

VOTE: Vision-Language-Action Optimization with Trajectory Ensemble Voting

On the Efficiency of LoRA Fine-Tuning for Vision-Language-Action Models in Industrial Robotic Manipulation