pith. sign in

arxiv: 2606.21470 · v1 · pith:3BU76VNGnew · submitted 2026-06-19 · 💻 cs.RO · cs.CV· cs.LG

ASCII Art Turns LLMs into VLA Controllers

Yitao Jiang , Roy Xing , Luyang Zhao , Brian Plancher , Muhao Chen , Devin Balkcom This is my paper

Pith reviewed 2026-06-26 14:30 UTC · model grok-4.3

classification 💻 cs.RO cs.CVcs.LG
keywords ASCII artvision-language-actionlarge language modelsrobot manipulationcontroller adaptationtext-only models
0
0 comments X

The pith

A text-only LLM can serve as a VLA controller when images are rendered as ASCII art.

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

The paper establishes that visual observations can be turned into text via ASCII art so that standard large language models can receive visual state, follow language instructions, and output robot actions. This removes the need to extend VLMs with action heads or train large multimodal models from scratch. Instead, existing LLM fine-tuning pipelines and deployment tools can be reused directly. The approach is tested on a 2D manipulation benchmark in both simulation and on physical hardware, where the controllers identify relevant objects and generate feasible action sequences.

Core claim

Rendering visual observations as ASCII art creates an interface that lets text-only LLMs condition on visual state, follow natural-language instructions, and produce constrained executable actions, enabling fine-tuning into VLA-style controllers without multimodal backbones.

What carries the argument

The ASCII-as-vision interface that converts images into text representations for direct input to LLMs.

If this is right

  • Existing LLM training stacks can be used to adapt models for visual control tasks without new multimodal infrastructure.
  • Controllers can process natural language instructions and output executable actions in both simulated and physical 2D manipulation settings.
  • The method complements conventional VLA pipelines by providing a lightweight modality bridge from images to text.
  • Iterative improvement is possible using techniques such as DAgger on top of expert demonstrations.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same ASCII bridge might allow text-only models to handle other sensor data types that can be serialized into text.
  • If ASCII representations prove sufficient for 2D tasks, they could be tested on tasks with higher spatial precision to measure the resolution limit.
  • Combining ASCII input with existing LLM safety or constraint mechanisms could directly constrain robot actions without additional vision modules.

Load-bearing premise

ASCII rendering preserves enough spatial and object information for manipulation tasks without critical loss that would block feasible action planning.

What would settle it

A manipulation task in which critical details such as exact object positions or shapes are lost in the ASCII rendering, causing the model to output invalid action sequences.

Figures

Figures reproduced from arXiv: 2606.21470 by Brian Plancher, Devin Balkcom, Luyang Zhao, Muhao Chen, Roy Xing, Yitao Jiang.

Figure 1
Figure 1. Figure 1: End-to-end pipeline showcasing training and testing in simulation [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: (a) Example of a generated Complex training scene (seed 26), showing steps 0–3 of a 13-step episode. The workspace contains three-colored objects and targets, obstacles, and a white circular gripper. The second row shows the ASCII encoding result. The third row was presented for human readability. (b) The ASCII encoding palette for our ASCII encoder. It is encoded in the HSV color space and has four edge c… view at source ↗
Figure 3
Figure 3. Figure 3: Real-robot image masking and denoising prior to ASCII encoding. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Reachability regimes and controlled scene generation. We first [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Stage-wise success rates in simulation. The 2 [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Real-robot pick-and-place rollout under the instruction “Pick the red object and place it at the red target.” The first row shows an external camera [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
read the original abstract

Vision--Language--Action (VLA) controllers are often built by extending vision--language models (VLMs) with action supervision, relying on multimodal backbones with large data and compute requirements. We demonstrate that a text-only large language model (LLM) can be adapted into a VLA-style controller when visual observations are rendered into a text input using an ASCII representation. This ASCII-as-vision interface enables existing training and deployment stacks for LLMs to efficiently condition on visual state, follow natural-language instructions, and produce constrained, executable actions. We fine-tune and compare multiple LLMs and VLMs across model families and scales, using both expert demonstrations from a planning-based teacher, as well as DAgger for iterative improvement. In a 2D manipulation benchmark, in both simulation and on a physical manipulator, the resulting controllers can identify task-relevant entities and plan feasible action sequences. Our results suggest that ASCII rendering can serve as a lightweight, interpretable modality bridge from images to text, complementing conventional VLA pipelines, and opening directions for VLA research with text-only backbones.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript claims that rendering visual observations as ASCII art enables fine-tuning of text-only LLMs into VLA-style controllers capable of following language instructions, identifying task-relevant entities, and generating executable actions for 2D manipulation tasks. It reports results using expert demonstrations and DAgger on a benchmark, with evaluation in both simulation and on a physical manipulator, positioning ASCII rendering as a lightweight modality bridge that avoids the need for multimodal VLM backbones.

Significance. If substantiated, the result would demonstrate a practical, low-overhead route for adapting existing LLM infrastructure to visual control problems, potentially lowering barriers to VLA research and offering an interpretable alternative to pixel-based or VLM-based pipelines. The approach could complement rather than replace conventional methods if it scales beyond the reported 2D setting.

major comments (3)
  1. [Abstract] Abstract: the central empirical claim of successful fine-tuning and real-robot deployment is stated without any quantitative metrics (success rates, trajectory errors, or comparisons), error bars, ablation results, or failure-case analysis, rendering the performance claims unverifiable from the provided text.
  2. [Abstract] Abstract and experimental description: no information-theoretic analysis, resolution study, or ablation on ASCII rendering parameters (grid size, character set, or preprocessing) is reported, which is load-bearing because the method's viability rests on the untested assumption that the lossy ASCII representation preserves sufficient spatial layout and object identity for feasible planning.
  3. [Experimental section] Experimental section: the use of DAgger for iterative improvement and comparison across model families is described at a high level, but without details on dataset sizes, training hyperparameters, action-space constraints, or how failures in entity identification were measured, it is impossible to assess whether the reported success stems from the ASCII interface or from the simplicity of the 2D benchmark scenes.
minor comments (2)
  1. [Abstract] The abstract and title use the term "VLA controllers" without an explicit definition or comparison to standard VLA formulations in the literature, which could be clarified for readers unfamiliar with the subfield.
  2. No mention of reproducibility artifacts (code, rendering scripts, or benchmark environments) is present in the provided text, which would strengthen the contribution if added.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback, which highlights opportunities to improve verifiability and detail in the manuscript. We address each major comment below, indicating revisions where the manuscript will be updated.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central empirical claim of successful fine-tuning and real-robot deployment is stated without any quantitative metrics (success rates, trajectory errors, or comparisons), error bars, ablation results, or failure-case analysis, rendering the performance claims unverifiable from the provided text.

    Authors: We agree the abstract would benefit from explicit metrics. The revised abstract now includes key quantitative results from our experiments (success rates in simulation and on hardware, with comparisons to baselines), along with pointers to ablations and failure-case analysis in the main text and supplementary material. This makes the central claims verifiable without exceeding length constraints. revision: yes

  2. Referee: [Abstract] Abstract and experimental description: no information-theoretic analysis, resolution study, or ablation on ASCII rendering parameters (grid size, character set, or preprocessing) is reported, which is load-bearing because the method's viability rests on the untested assumption that the lossy ASCII representation preserves sufficient spatial layout and object identity for feasible planning.

    Authors: The original manuscript provides empirical validation through task performance but lacks systematic parameter ablations. We have added a dedicated subsection with resolution studies and ablations on grid size, character set, and preprocessing, demonstrating their effects on entity identification and planning success. While a formal information-theoretic analysis is not included (due to the discrete, non-probabilistic nature of ASCII rendering), the empirical results across varied parameters substantiate that sufficient spatial and identity information is preserved for the 2D tasks considered. revision: partial

  3. Referee: [Experimental section] Experimental section: the use of DAgger for iterative improvement and comparison across model families is described at a high level, but without details on dataset sizes, training hyperparameters, action-space constraints, or how failures in entity identification were measured, it is impossible to assess whether the reported success stems from the ASCII interface or from the simplicity of the 2D benchmark scenes.

    Authors: We have expanded the experimental section with the requested details: dataset sizes (number of expert demonstrations and DAgger iterations), full training hyperparameters, explicit action-space constraints, and the protocol for measuring entity identification failures (via per-step logging against ground-truth object positions). These additions clarify the contribution of the ASCII interface and allow readers to evaluate generalization beyond the benchmark simplicity. revision: yes

Circularity Check

0 steps flagged

No circularity; empirical demonstration rests on experiments, not derivations or self-referential fits

full rationale

The paper presents an empirical adaptation of text-only LLMs to VLA-style control via ASCII rendering of visual observations, validated through fine-tuning on expert demonstrations and DAgger, then evaluated on a 2D manipulation benchmark in simulation and on hardware. No equations, mathematical derivations, or parameter-fitting steps are described that could reduce a claimed prediction to its inputs by construction. No self-citations are invoked as load-bearing uniqueness theorems or ansatzes. The central claim is supported by reported experimental outcomes rather than any self-definitional or fitted-input logic, making the work self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review; no explicit free parameters, axioms, or invented entities are stated. Standard supervised fine-tuning assumptions and the sufficiency of ASCII for 2D scenes are implicit but unexamined.

pith-pipeline@v0.9.1-grok · 5735 in / 998 out tokens · 20651 ms · 2026-06-26T14:30:38.914382+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

26 extracted references · 14 canonical work pages · 8 internal anchors

  1. [1]

    RT-1: Robotics Transformer for Real-World Control at Scale,

    A. Brohanet al., “RT-1: Robotics Transformer for Real-World Control at Scale,” inRobotics: Science and Systems XIX. Robotics: Science and Systems Foundation, Jul 2023

    2023

  2. [2]

    NanoVLA: Routing Decoupled Vision-Language Un- derstanding for Nano-sized Generalist Robotic Policies,

    J. Chenet al., “NanoVLA: Routing Decoupled Vision-Language Un- derstanding for Nano-sized Generalist Robotic Policies,” Oct 2025, arXiv:2510.25122

    work page arXiv 2025

  3. [3]

    Reducing the Barrier to Entry of Complex Robotic Software: a MoveIt! Case Study

    D. Colemanet al., “Reducing the Barrier to Entry of Complex Robotic Software: a MoveIt! Case Study,” Apr 2014, arXiv:1404.3785

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  4. [4]

    Open X-Embodiment: Robotic Learning Datasets and RT-X Models

    O. X.-E. Collaborationet al., “Open X-Embodiment: Robotic Learning Datasets and RT-X Models,” May 2025, arXiv:2310.08864

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  5. [5]

    PaLM-E: An Embodied Multimodal Language Model,

    D. Driesset al., “PaLM-E: An Embodied Multimodal Language Model,” inProceedings of the 40th International Conference on Machine Learn- ing. PMLR, Jul 2023, pp. 8469–8488, ISSN: 2640-3498

    2023

  6. [6]

    Octo: An Open-Source Generalist Robot Policy,

    D. Ghoshet al., “Octo: An Open-Source Generalist Robot Policy,” in Robotics: Science and Systems XX. Robotics: Science and Systems Foundation, Jul 2024

    2024

  7. [7]

    Good old-fashioned engineering can close the 100,000- year “data gap

    K. Goldberg, “Good old-fashioned engineering can close the 100,000- year “data gap” in robotics,”Science Robotics, vol. 10, no. 105, p. eaea7390, Aug 2025, publisher: American Association for the Advance- ment of Science

    2025

  8. [8]

    A formal basis for the heuristic determination of minimum cost paths,

    P. E. Hart, N. J. Nilsson, and B. Raphael, “A formal basis for the heuristic determination of minimum cost paths,”IEEE Transactions on Systems Science and Cybernetics, vol. 4, no. 2, pp. 100–107, 1968

    1968

  9. [9]

    LoRA: Low-Rank Adaptation of Large Language Models

    E. J. Huet al., “LoRA: Low-Rank Adaptation of Large Language Models,” Oct 2021, arXiv:2106.09685

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  10. [10]

    VoxPoser: Composable 3D Value Maps for Robotic Manipulation with Language Models

    W. Huanget al., “V oxPoser: Composable 3D Value Maps for Robotic Manipulation with Language Models,” Nov 2023, arXiv:2307.05973

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  11. [11]

    ASCIIEval: Benchmarking Models’ Visual Perception in Text Strings via ASCII Art,

    Q. Jiaet al., “ASCIIEval: Benchmarking Models’ Visual Perception in Text Strings via ASCII Art,” Sep 2025, arXiv:2410.01733 [cs]

    work page arXiv 2025

  12. [12]

    Exploring Spontaneous Social Interaction Swarm Robotics Powered by Large Language Models,

    Y . Jianget al., “Exploring Spontaneous Social Interaction Swarm Robotics Powered by Large Language Models,” in2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Oct 2025, pp. 10 198–10 205, ISSN: 2153-0866

    2025

  13. [13]

    VIMA: Robot Manipulation with Multimodal Prompts,

    ——, “VIMA: Robot Manipulation with Multimodal Prompts,” inPro- ceedings of the 40th International Conference on Machine Learning. PMLR, Jul 2023, pp. 14 975–15 022, ISSN: 2640-3498

    2023

  14. [14]

    Vision-Language-Action Models for Robotics: A Review Towards Real-World Applications,

    K. Kawaharazukaet al., “Vision-Language-Action Models for Robotics: A Review Towards Real-World Applications,”IEEE Access, vol. 13, pp. 162 467–162 504, 2025, arXiv:2510.07077

    work page arXiv 2025

  15. [15]

    OpenVLA: An Open-Source Vision-Language-Action Model

    M. J. Kimet al., “OpenVLA: An Open-Source Vision-Language-Action Model,” Sep 2024, arXiv:2406.09246

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  16. [16]

    Optimization-based locomotion planning, esti- mation, and control design for the atlas humanoid robot,

    S. Kuindersmaet al., “Optimization-based locomotion planning, esti- mation, and control design for the atlas humanoid robot,”Autonomous Robots, vol. 40, no. 3, pp. 429–455, Mar 2016

    2016

  17. [17]

    Code as Policies: Language Model Programs for Embodied Control

    J. Lianget al., “Code as Policies: Language Model Programs for Embodied Control,” May 2023, arXiv:2209.07753

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  18. [18]

    ASCIIBench: Evaluating Language-Model-Based Under- standing of Visually-Oriented Text,

    K. Luoet al., “ASCIIBench: Evaluating Language-Model-Based Under- standing of Visually-Oriented Text,” Dec 2025, arXiv:2512.04125 [cs]

    work page arXiv 2025

  19. [19]

    GRID: Scene-Graph-based Instruction-driven Robotic Task Planning,

    Z. Niet al., “GRID: Scene-Graph-based Instruction-driven Robotic Task Planning,” in2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Oct 2024, pp. 13 765–13 772, ISSN: 2153- 0866

    2024

  20. [20]

    SayPlan: Grounding Large Language Models using 3D Scene Graphs for Scalable Robot Task Planning,

    K. Ranaet al., “SayPlan: Grounding Large Language Models using 3D Scene Graphs for Scalable Robot Task Planning,” Sep 2023, arXiv:2307.06135

    work page arXiv 2023

  21. [21]

    Flower: Democratizing generalist robot policies with efficient vision-language-action flow policies,

    M. Reusset al., “FLOWER: Democratizing Generalist Robot Poli- cies with Efficient Vision-Language-Action Flow Policies,” Sep 2025, arXiv:2509.04996

    work page arXiv 2025

  22. [22]

    A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning,

    S. Ross, G. Gordon, and D. Bagnell, “A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning,” inProceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics. JMLR Workshop and Conference Proceedings, Jun 2011, pp. 627–635, ISSN: 1938-7228. [Online]. Available: https://proceedings.mlr.pr...

    2011

  23. [23]

    SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics

    M. Shukoret al., “SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics,” Jun 2025, arXiv:2506.01844

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  24. [24]

    CERBERUS in the DARPA Subterranean Chal- lenge,

    M. Tranzattoet al., “CERBERUS in the DARPA Subterranean Chal- lenge,”Science Robotics, vol. 7, no. 66, p. eabp9742, May 2022, publisher: American Association for the Advancement of Science

    2022

  25. [25]

    CoT-VLA: Visual Chain-of-Thought Reasoning for Vision-Language-Action Models

    Q. Zhaoet al., “CoT-VLA: Visual Chain-of-Thought Reasoning for Vision-Language-Action Models,” Mar 2025, arXiv:2503.22020

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  26. [26]

    RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control,

    B. Zitkovichet al., “RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control,” inProceedings of The 7th Con- ference on Robot Learning. PMLR, Dec 2023, pp. 2165–2183, ISSN: 2640-3498

    2023