Skip to content

PixelVLA: Advancing Pixel-level Understanding in Vision-Language-Action Model

Conference: ICLR 2026
OpenReview: https://openreview.net/forum?id=7M6ryCABIc
Paper: Project Page
Code: https://wenqiliang.github.io/PixelVLA/ (Project Page)
Area: Robotics / Embodied AI / VLA
Keywords: Vision-Language-Action Models, Pixel-level Understanding, Visual Prompting, Continuous Actions, Visuomotor Instruction Tuning

TL;DR

PixelVLA is the first vision-language-action model to support both pixel-level understanding and multi-modal prompting (text + points/lines/boxes/masks). By integrating three components—a "multi-scale pixel-aware encoder, a visual prompt encoder, and a continuous action decoder"—into existing VLAs and utilizing an automated annotation pipeline to create the Pixel-160K dataset, it enhances manipulation success rates by \(10.1\% \sim 28.7\%\) over OpenVLA at only \(1.5\%\) of the pre-training cost.

Background & Motivation

Background: Vision-Language-Action (VLA) models combine large-scale robotic data with pre-trained Vision-Language Models (VLMs) to generalize to unseen objects and instructions in zero-shot scenarios. Representative works like RT-2, OpenVLA, and \(\pi_0\) have become the mainstream paradigm for general-purpose robotic manipulation policies.

Limitations of Prior Work: Almost all existing VLAs directly inherit from VLMs, which process observations at the "image-level." While they can identify an eggplant in a scene, they lack precision regarding its exact pixel contours and spatial boundaries. This leads to two specific issues: first, a lack of fine-grained scene understanding and weak spatial reasoning, resulting in poor out-of-distribution generalization; second, most VLAs only accept text instructions, ignoring intuitive visual cues like points, lines, boxes, or masks. This limits human-robot interaction, particularly for referential instructions like "put this eggplant into the basket."

Key Challenge: While pixel-level understanding has been validated in VLMs (e.g., Shikra, Ferret), migrating it to VLAs is hindered by data bottlenecks. Existing robot datasets (OXE, DROID, etc.) lack multi-modal visual prompts and pixel-level mask annotations. Directly using off-the-shelf open-vocabulary segmentation models to label robotic images yields poor results due to cluttered, low-quality images and a significant domain gap with VLM pre-training data.

Goal: (1) Design a model architecture that enables VLAs to gain pixel-level understanding and accept multi-modal visual prompts; (2) Solve the "lack of pixel annotations in robotic data" bottleneck by automatically generating usable training data; (3) Design a training pipeline to inject these capabilities into existing VLAs at low cost.

Key Insight: The authors propose Visuomotor Instruction Tuning, migrating the mature "visual instruction tuning" from VLMs to VLAs. This expands action generation from \(a_t = F_\theta(x_t, L)\) (image + text \(\to\) action) to \(a_t = F_\theta(x_t, p_t, L, V)\), introducing pixel mask inputs \(p_t\) and diverse visual prompts \(V\).

Core Idea: Encode pixel-level information from segmentation masks into LLM tokens via "pixel-aware embeddings." A lightweight visual prompt encoder receives points/lines/boxes/masks, and a continuous action decoder replaces discrete action tokens. These three modules enhance existing VLAs in a plug-and-play manner, supported by an automated annotation pipeline to fill the missing pixel-level training data.

Method

Overall Architecture

PixelVLA enables VLAs to perceive pixels and follow visual prompts. It utilizes the Prismatic-7B backbone (DinoV2 + SigLIP encoders + Llama 2-7B). Three new components are inserted: Observations, language instructions, and visual prompts are processed through the visual encoder for embeddings; the multi-scale pixel-aware encoder aggregates multi-layer features from mask regions into pixel-aware embeddings; and the visual prompt encoder encodes points/lines/boxes/masks into prompt embeddings. These are fed into the LLM, and the hidden state of the final layer is used by the continuous action decoder to regress 7-dimensional robotic actions (incorporating action chunking).

This structure is supported by a two-stage automated labeling pipeline that processes public datasets (Fractal, Bridge v2) into the Pixel-160K dataset, and two-stage visuomotor instruction tuning, which first learns continuous action representations and then injects pixel-level understanding via LoRA.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Input: Robot Observation<br/>+ Language Instruction + Visual Prompt"] --> P["Two-stage Automated Labeling Pipeline<br/>→ Pixel-160K Dataset"]
    A --> V["Visual Encoder<br/>DinoV2 + SigLIP + MLP"]
    P -->|Provide Mask & Prompt Labels| ENC
    subgraph ENC["Pixel + Prompt Dual Encoding"]
        direction TB
        M["Multi-scale Pixel-aware Encoder<br/>Mask-weighted Multi-layer Features → Pixel Embeddings"]
        S["Visual Prompt Encoder<br/>Point/Line/Box/Mask → Prompt Embeddings"]
    end
    V --> L["LLM Backbone<br/>Llama 2-7B"]
    ENC --> L
    L --> D["Continuous Action Decoder<br/>Hidden State → Continuous Actions"]
    D --> O["Output: 7D Robot Action"]

Key Designs

1. Multi-scale Pixel-aware Encoder: Compressing mask region features into LLM-readable pixel embeddings

This addresses the "pixel blindness" of VLAs. Given an observation \(x_0\), SigLIP extracts multi-layer visual features \(F_v^0=\{f_v^{0,i}\}_{i=1}^L\). The pixel mask \(p_0\) defines the target region. The core mechanism performs mask-weighted averaging on each feature layer, followed by linear projection, cross-layer summation, and an MLP to obtain the pixel-aware embedding \(E_p^0\):

\[E_p^0 = \mathrm{MLP}\Big(\sum_{i=1}^{L}\Gamma_i(f_p^{0,i})\Big),\qquad f_p^{0,i}=\frac{p_0\cdot f_v^{0,i}}{|p_0|}\]

Where \(\Gamma_i(\cdot)\) is the linear projection for the \(i\)-th layer, and \(f_p^{0,i}\) is the normalized average feature for that region. Since these embeddings are supervised by action prediction loss, the model learns the "pixel info \(\leftrightarrow\) action" association, truly injecting pixel-level understanding into the VLA backbone.

2. Visual Prompt Encoder: Unified processing of points/lines/boxes/masks

This enables VLAs to receive visual instructions. Borrowing from SAM’s prompt encoder, input visual prompts \(V_0\) are converted to continuous positional embeddings based on normalized coordinates, overlaid with learnable "prompt type embeddings" to form feature \(F_s^0\), and processed via MLP into \(E_s^0\). Explicit spatial binding ensures the LLM retains precise location information, allowing users to point to objects instead of providing verbose text descriptions.

3. Continuous Action Decoder: Direct regression to bypass discretization loss

Existing models like OpenVLA discretize each action dimension into 256 bins, which loses fine-grained detail. PixelVLA follows the continuous approach of \(\pi_0\): the final LLM hidden state \(F_t\) passes through a linear projection, \(N_r\) ResNet blocks, and an MLP to output continuous actions \(A\in\mathbb{R}^{N_c\times 7}\), where \(N_c\) is the action chunking size. This combination avoids cumulative discretization errors and captures longer temporal dependencies, significantly improving performance on long-horizon tasks like LIBERO-Long.

4. Two-stage Automated Labeling Pipeline \(\to\) Pixel-160K: Extracting pixel labels from robotic videos

The pipeline consists of: Gripper-aware Region Proposal—identifying the "gripper closure" moment as a heuristic for the object's location. SAM 2 segments the gripper mask in the closure frame, and an expanded bounding box defines the region proposal \(R_\eta\). Multi-modal Object Segmentation—using Llama 2-7B to extract the target noun (e.g., "Eggplant") from the instruction and feeding it with \(R_\eta\) into Grounding DINO and SAM. After filtering by confidence, prompts (points/lines/boxes) are randomly sampled from the resulting masks. The final Pixel-160K dataset contains 160k episodes and 6.5 million visual-prompt-mask-action triplets.

Loss & Training

A two-stage visuomotor instruction tuning strategy is used. Stage 1 (Continuous Action Training, CAT): The LLM and encoders are initialized with weights from OpenVLA/\(\pi_0\). Except for the continuous action decoder, most modules are frozen. Only the \(L_1\) regression loss is used to align LLM hidden states with ground-truth continuous actions on Fractal + Bridge v2 data. Stage 2 (Pixel-level Understanding Enhancement, PUE): The LLM backbone is fine-tuned using LoRA (\(r=32\)) on Pixel-160K, while jointly training the prompt and pixel encoders and optimizing the action decoder. The loss is:

\[\mathcal{L}_{PixelVLA}=\sum_{i=1}^{B}\big\|a_i - C\big(H(E_v^i, E_l^i, E_p^i, E_s^i)\big)\big\|_1\]

Where \(C(\cdot)\) is the decoder, \(H\) is the LLM, and \(E_v/E_l/E_p/E_s\) are the diverse embeddings.

Key Experimental Results

Evaluated on SimplerEnv-Google Robot, SimplerEnv-WidowX, and LIBERO.

Main Results

Benchmark / Setting Metric PixelVLA Baseline Gain
Google Robot (VM Avg) Success Rate 61.4 OpenVLA 32.7 +28.7
Google Robot (VA Avg) Success Rate 50.1 OpenVLA 40.0 +10.1
WidowX (Suc. Avg, \(\pi_0\) base) Success Rate 33.8 \(\pi_0\) 27.1 +6.7
LIBERO (Avg of 4 sets) Success Rate 86.7 OpenVLA 76.5 +10.2
LIBERO-Long Success Rate 82.6 OpenVLA 53.7 +28.9

PixelVLA achieves SOTA on LIBERO and leads significantly on LIBERO-Long. This is attributed to the continuous decoder and action chunking, which capture temporal dependencies and mitigate discretization errors.

Ablation Study

Ablation on Google Robot (VA) using OpenVLA as the baseline:

Configuration Avg Success Rate Description
Baseline 40.0 Original OpenVLA
+FT 37.0 FT on Fractal+Bridge only (decreases)
+FT+CAT 43.8 With Continuous Action Training, +3.8
+FT+PUE 48.0 With Pixel-level Enhancement, +8.0
PixelVLA (CAT+PUE) 50.1 Full two-stage, +6.3 over CAT

Key Findings

  • PUE is the primary driver: Adding PUE alone (+8.0) significantly outperforms adding CAT alone (+3.8), proving pixel-level info is key for generalization.
  • Direct fine-tuning can be detrimental: Pure fine-tuning (+FT) scored lower than the baseline, indicating improvements stem from the new components and two-stage paradigm rather than just "more data."
  • Low Cost: These results were achieved using only \(\sim 1.5\%\) of OpenVLA's pre-training compute.

Highlights & Insights

  • "Gripper closure" as a positioning heuristic: Positioning target objects via gripper closure moments simplifies pixel annotation into an automated pipeline, avoiding the need for per-frame manual labeling.
  • Implicit supervision through action loss: Pixel-aware embeddings are learned via downstream action regression rather than a dedicated segmentation loss, simplifying multi-task optimization.
  • Plug-and-play: The architecture shows consistent gains on both OpenVLA and \(\pi_0\), demonstrating versatility across different VLA backbones.

Limitations & Future Work

  • Simulation-only evaluation: Results are currently limited to SimplerEnv and LIBERO; real-robot deployment and sim-to-real performance remain to be verified.
  • Resolution and perspective: Dependency on a single third-person view and \(224 \times 224\) resolution may affect robustness in high-resolution or occluded scenarios.
  • Catastrophic forgetting: Two-stage joint optimization showed slight performance drops in highly sensitive tasks (e.g., Open/Close Drawer), requiring better mitigation strategies.
  • vs OpenVLA: PixelVLA replaces discrete tokens with continuous actions and adds pixel/prompt encoders, surpassing it by \(10.1\% \sim 28.7\%\) at minimal cost.
  • vs TraceVLA: TraceVLA uses visual trajectory prompts for spatio-temporal awareness; PixelVLA reaches pixel-mask level granularity, outperforming it on Google Robot and LIBERO.
  • vs Region-level VLMs (Shikra, Ferret): While those focus on perception, PixelVLA is the first to align fine-grained pixel masks with continuous actions for embodied control.

Rating

  • Novelty: ⭐⭐⭐⭐⭐ First to implement both pixel-level understanding and multi-modal prompting in VLA.
  • Experimental Thoroughness: ⭐⭐⭐⭐ Solid benchmark results and ablations, though lacks real-world testing.
  • Writing Quality: ⭐⭐⭐⭐ Clear structure and well-defined mechanisms.
  • Value: ⭐⭐⭐⭐⭐ Plug-and-play, low-cost enhancement of VLA fine-grained perception.