Evo-1: Lightweight Vision-Language-Action Model with Preserved Semantic Alignment

Conference: CVPR 2026
Paper: CVF Open Access
Code: https://github.com/MINT-SJTU/Evo-1 (Available)
Area: Embodied AI / Vision-Language-Action (VLA)
Keywords: Lightweight VLA, Semantic Alignment Preservation, Flow-matching Diffusion, Cross-attention DiT, Two-stage Training

TL;DR

Evo-1 utilizes a native multimodal VLM with only 0.77B parameters as the backbone, paired with a pure cross-attention flow-matching diffusion action expert and a "freeze-then-fine-tune" two-stage training strategy. Without any robot data pre-training, it achieves SOTA on Meta-World, RoboTwin, and LIBERO by preserving the VLM's semantic space, reaching a 78% success rate in real-world tests with 16.4 Hz inference and only 2.3 GB VRAM.

Background & Motivation

Background: Vision-Language-Action (VLA) models unify perception, language, and control into a multimodal framework, enabling robots to "see images, hear instructions, and perform actions." Mainstream approaches (OpenVLA, \(\pi_0\), GR00T, etc.) typically employ large VLMs with billions of parameters as backbones and undergo extensive pre-training on large-scale robot datasets like OXE/DROID to gain strong generalization.

Limitations of Prior Work: This path faces four specific issues: ① Billion-parameter scales consume massive VRAM and compute; ② High computational load leads to low control frequency and slow real-world reaction; ③ Standard end-to-end joint training tends to destroy the representation space of the VLM backbone, leading to downstream overfitting and poor generalization; ④ Heavy reliance on expensive and labor-intensive large-scale robot data pre-training. Existing lightweight solutions (TinyVLA, SmolVLA) reduce parameters but lack performance and robustness in complex manipulation tasks.

Key Challenge: There is an overlooked tension between preserving pre-trained VLM multimodal semantics and adapting to downstream action generation. Direct end-to-end joint training allows noisy gradients from randomly initialized action heads to backpropagate into the VLM, disrupting the pre-aligned vision-language attention (the paper illustrates this "semantic drift" via attention maps).

Goal: Achieve high success rates and inference frequency under the constraints of no robot data pre-training and < 1B parameters, while preserving the backbone's generalization capability.

Key Insight: Rather than using a stitched backbone (pure-text LLM with a post-hoc vision aligner), it is more effective to use a native multimodal pre-trained compact VLM (InternVL3-1B), where vision-language representations are already tightly aligned. Furthermore, a training schedule is needed to prevent the action head from "polluting" the backbone.

Core Idea: A tripartite system consisting of a "native multimodal lightweight backbone + pure cross-attention flow-matching action expert + two-stage (freeze → fine-tune) training." By prioritizing "semantic alignment" as a protected asset, the model matches or exceeds large models despite small parameters and zero robot pre-training.

Method

Overall Architecture

Evo-1 is a modular VLA: given multi-view RGB observations \(\{I_t^i\}_{i=1}^N\), language instructions \(L_t\), and robot proprioception \(s_t\), it outputs a continuous action vector \(a_t \in \mathbb{R}^{d_a}\), formulated as \(a_t = f_{\text{Evo-1}}(\{I_t^i\}_{i=1}^N, L_t, s_t; \theta)\). It consists of three core components: ① A VLM backbone encoding images and instructions into a fused representation \(z_t\); ② An integration module aligning and concatenating \(z_t\) with \(s_t\) for the controller; ③ A Cross-modulated Diffusion Transformer (action expert) generating a sequence of future actions via flow-matching under these conditions. A two-stage training schedule determines which parts are frozen or thawed.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Multi-view RGB + Instruction + Proprioception"] --> B["VLM Backbone<br/>Native Multimodal Lightweight Encoding<br/>Extract 14th Layer Fused Feature z_t"]
    B --> C["Integration Module<br/>z_t and s_t Concat as KV"]
    C --> D["Cross-modulated DiT<br/>Pure Cross-attention + Flow-matching"]
    D --> E["Continuous Action Chunk â_t..â_t+H-1"]
    F["Two-stage Training<br/>Freeze Backbone → Full Fine-tuning"] -.Schedule.-> B
    F -.Schedule.-> C
    F -.Schedule.-> D

Key Designs

1. Native Multimodal Lightweight Backbone: Replacing Stitched Backbones with Tightly Aligned Small VLMs

Addressing limitations ①② (large parameters, low frequency) and ③ (representation destruction), Evo-1 selects InternVL3-1B as the backbone instead of stitched 7B backbones like OpenVLA. InternVL3 learns vision and language jointly under a single-stage native multimodal paradigm, resulting in better cross-modal alignment. Consequently, after downstream training, its attention maps (Fig. 2) maintain spatial consistency and semantic focus, while Prismatic-7B exhibits significant drift. Specifically, it uses InternViT-300M (distilled from InternViT-6B), where RGB images are scaled to 448×448 and pixel-unshuffled to reduce vision tokens to 1/4, providing compact yet spatially granular patch embeddings. The language side uses Qwen2.5-0.5B. During fusion, patch-level image embeddings replace <img> placeholder tokens in the sequence, passing through a shared decoder to obtain \(z_t = f_{\text{VLM}}(\{I_t^i\}, L_t)\). A key trade-off is keeping only the first 14 layers of the language branch—the middle layers were found to have the strongest cross-modal alignment and are most useful for visual motor control; removing deeper layers saves compute without losing alignment info.

2. Cross-modulated Diffusion Transformer: Pure Cross-attention Flow-matching Action Expert

To efficiently generate coherent continuous actions, Evo-1's action expert is a DiT composed only of stacked cross-attention layers, intentionally omitting the alternating self-attention and cross-attention structure used in \(\pi_0\)/SmolVLA. The authors demonstrate that such alternation can interrupt the continuous propagation of multimodal information. It follows the flow-matching paradigm, learning a time-dependent velocity field to gradually push initial noise toward the ground-truth action. During training, ground-truth actions \(A_t\) and random noise \(\epsilon\) are linearly interpolated:

\[A_t^\tau = \tau A_t + (1-\tau)\epsilon,\]

where interpolation weight \(\tau\) is sampled from a Beta distribution and clipped to \([0.02, 0.98]\) for stability. The action expert learns a velocity field \(v_\theta\) conditioned on \(z_t\) and \(s_t\), with the objective:

\[\mathcal{L}^\tau(\theta) = \mathbb{E}\left[\,\left\| v_\theta(A_t^\tau, z_t, s_t) - u(A_t^\tau \mid A_t) \right\|^2\,\right],\]

where \(u(A_t^\tau \mid A_t)\) is the target flow direction. At inference, it predicts an action chunk of length \(H\): \(\hat A_t = f_{\text{AE}}(z_t, s_t, A_t^\tau)\). The pure cross-attention and action chunking enable a compact structure and high inference frequency (16.4 Hz).

3. Integration Module: "Concatenation" vs. "Projection" of Mid-layer Features and Proprioception

To prevent loss of perceptual information, the integration module extracts fused features \(z_t\) from the 14th layer of the VLM (mid-layer semantics balancing vision and language). It then directly concatenates \(z_t\) with proprioception \(s_t\) rather than projecting them into a shared embedding space, preserving the integrity of both. The combined features serve as the key-value input for all DiT layers in the action expert, while noisy actions \(A_t^\tau\) act as the query. Ablations compare this (Module A: Mid-Layer Cross-Attention) with three variants: B (alternating self/cross-attention), C (layer-wise injection of VLM features), and D (concatenating all as joint KV). Module A is optimal as it provides the most continuous multimodal information propagation.

4. Two-stage Training: Freeze Backbone then Full Fine-tune to Protect Semantic Space

This is the core design addressing the destruction of representations. Direct joint training forces gradients from the random action head into the VLM, distorting pre-trained semantics and causing overfitting. Evo-1 splits training: Stage 1 (Action Expert Alignment)—Freeze the entire VLM backbone and train only the integration module and action expert, allowing the action head to align with the multimodal embedding space without polluting the backbone. Stage 2 (Full Fine-tuning)—Thaw the VLM after alignment stabilizes, jointly fine-tuning the entire architecture to deeply couple the backbone and action head. Attention visualizations (Fig. 7) show that after two-stage training, attention remains focused on task-relevant regions, whereas single-stage training leads to attention scattering.

Loss & Training

The core training objective is the flow-matching velocity regression loss \(\mathcal{L}^\tau(\theta)\). The training follows the two-stage schedule. Simulation tasks use ~50 demonstrations per task, and real-world tasks use 100 teleoperated demonstrations, without any large-scale robot pre-training.

Key Experimental Results

Main Results

Simulation benchmarks (Success Rate %, higher is better; Evo-1 0.77B without robot pre-training):

Benchmark	Metric	Evo-1 (0.77B)	Prev. SOTA	Gain
Meta-World	Average SR	80.6	SmolVLA 68.2 (2.25B)	+12.4
RoboTwin	Average SR	37.8	\(\pi_0\) 30.9 (3.5B)	+6.9
LIBERO	Average SR	94.8	\(\pi_0\) 94.2 (3.5B)	+0.6

On Meta-World, Evo-1 leads across all four difficulty levels; in "very hard," it achieves 79.2% (vs. SmolVLA 64.0%). In RoboTwin's "Click Alarmclock" (hard), it reaches 58% (vs. \(\pi_0\) 11%), showing strong bimanual coordination.

Real-world tasks (xArm6, 20 trials/task) + Inference efficiency (RTX 4090d):

Model	Params (B)	VRAM (GB)	Frequency (Hz)	Real-world SR (%)
SmolVLA	0.45	2.0	12.7	50.0
OpenVLA	7.0	15.1	7.9	55.0
\(\pi_0\)	3.5	17.9	11.5	73.0
Ours	0.77	2.3	16.4	78.0

Evo-1 outperforms \(\pi_0\) in VRAM, frequency, and success rate with approximately 1/4 of the parameters.

Ablation Study

Configuration	Validation Benchmark	Conclusion
Integration Module A (Ours)	LIBERO-Long	Optimal - continuous info propagation
Integration Module B (Alt. SA/CA)	LIBERO-Long	Self-attention interrupts propagation
Integration Module C (Layer-wise)	LIBERO-Long	Inconsistent layer conditions
Integration Module D (Joint KV)	LIBERO-Long	Cross-layer inconsistency
Two-stage Training (Ours)	Meta-World	Outperforms single-stage across all difficulties
Single-stage Training	Meta-World	Semantic drift, scattered attention

Key Findings

• Information propagation continuity is the key to integration module success: Module A wins because a consistent mid-layer feature + state is fed to all DiT layers.
• Two-stage training gains come from "preserving semantics": Attention maps show single-stage training causes the model to attend to irrelevant regions, whereas two-stage training preserves focus on task entities.
• Generalization Robustness: In real-world Pick-and-Place distractor experiments, Evo-1 outperforms SmolVLA with unseen distractors (80% vs. 65%), background color changes (75% vs. 60%), and target displacement.
• Small Backbone + Native Multimodal > Large Backbone + Stitched Alignment: InternVL3-1B produces more stable attention maps than Prismatic-7B, validating the value of "native alignment."

Highlights & Insights

• Treating "Semantic Alignment" as a protected first-class citizen: While most VLAs default to end-to-end training, this work freezes the backbone first and quantifies "semantic drift," making the training schedule itself a core contribution.
• Counter-intuitive choice of pure cross-attention DiT: Removing self-attention is better because it prevents interrupting the continuous propagation of multimodal conditions.
• Concatenation instead of Projection: Using concat for the integration module avoids information compression caused by shared-space projection.
• Engineering Friendly: 0.77B + 2.3GB VRAM + 16.4Hz makes real-time control feasible on consumer-grade GPUs.

Limitations & Future Work

• The paper does not provide fine-grained decoupling of design points (e.g., specific contributions of "native backbone" vs. "two-stage training").
• ⚠️ Real-world generalization tests were performed only on a single Pick-and-Place task; robustness to more dramatic distribution shifts remains to be verified.
• Hyperparameters like extracting the 14th layer are empirical; optimal layers might need to be re-searched for different backbones.
• Still relies on dozens to hundreds of demos per task; few-shot/zero-shot capability for entirely new tasks is not yet demonstrated.

Related Work & Insights

• vs. OpenVLA: OpenVLA uses 7B Prismatic + discrete action modeling + OXE pre-training; Evo-1 uses 0.77B native multimodal backbone + continuous flow-matching + zero robot pre-training, achieving higher success (78% vs. 55%) with an order of magnitude fewer parameters.
• vs. \(\pi_0\): \(\pi_0\) is based on PaliGemma (3.5B) + flow-matching with robot pre-training; Evo-1 uses a pure cross-attention DiT with ~1/4 the parameters and outperforms it in both simulation and real-world tests.
• vs. SmolVLA/TinyVLA: These share the sub-billion parameter scale but lack robustness in complex tasks; Evo-1 differentiates itself via the native alignment backbone and semantic-preserving two-stage training.

Rating

• Novelty: ⭐⭐⭐⭐ Combinatorial innovation (native backbone + pure cross-attention DiT + two-stage training) with a clear semantic focus.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ Comprehensive evaluation across three benchmarks, real-world tasks, and efficiency/generalization/ablations.
• Writing Quality: ⭐⭐⭐⭐ Clear structure; strong evidence via attention maps.
• Value: ⭐⭐⭐⭐⭐ SOTA results with 0.77B parameters and zero robot pre-training; highly deployable.

Related Papers

• [AAAI 2026] Continuous Vision-Language-Action Co-Learning with Semantic-Physical Alignment for Behavioral Cloning
• [ICLR 2026] Spatial Forcing: Implicit Spatial Representation Alignment for Vision-language-action Model
• [CVPR 2026] Mantis: A Versatile Vision-Language-Action Model with Disentangled Visual Foresight
• [CVPR 2026] MergeVLA: Cross-Skill Model Merging Toward a Generalist Vision-Language-Action Agent
• [CVPR 2026] QuantVLA: Scale-Calibrated Post-Training Quantization for Vision-Language-Action Models