DualVLA: Building a Generalizable Embodied Agent via Partial Decoupling of Reasoning and Action

• Specialist VLAs [23] [48] [74] achieve strong manipulation but limited multimodal reasoning; adding reasoning can reduce manipulation even with larger, more diverse training data.The paper names this decline action degeneration and suggests reasoning and action rely on shared internal representations.Source: §1 · p. 2
• Dual-layer pruning uses embodiment cues and scene-level event changes to remove redundant embodied reasoning while retaining action-critical content.The method addresses how repetitive, low-entropy reasoning can bias optimization and deteriorate action execution.Source: §1 · p. 2
• Dual-teacher adaptive distillation uses specialist-VLA action supervision and distinct soft labels for robot and multimodal reasoning data.This provides balanced, fine-grained guidance for learning manipulation and reasoning.Source: §1 · p. 2
• VLA Score uses a strong VLM evaluator to assess action, reasoning, intention, and reasoning–action alignment in reasoning-capable VLAs.The framework is designed to expose performance trade-offs that existing evaluations do not disentangle.Source: §1 · p. 2
• DUALVLA’s post-training framework targets action degeneration while preserving multimodal reasoning and improving manipulation.It decouples reasoning and action learning at the data and loss levels.Source: §1 · p. 2
• Experiments across simulation and real-world tasks show DUALVLA mitigates action degeneration and improves overall VLA performance.VLA Score also reveals bottlenecks in current VLA development and validates the approach’s strengths.Source: §1 · p. 2

Vision-Language-Action (VLA) Models

• Specialist VLAs fine-tune vision-language models on robot data for action, but reasoning VLAs add embodied chain-of-thought and multimodal data while showing degraded action performance.Emma-X uses Gemini-generated embodied CoT annotations to fine-tune OpenVLA; Table 1 compares base model, data, reasoning capability, and action capability across VLMs, specialist VLAs, and reasoning VLAs.Source: §2 · pp. 2–3, Table 1 · p. 3
• OpenVLA fine-tunes Prismatic-7B and maps discrete tokens to action bins, while later studies explore alternative architectures and action modeling [23] [21] [2] [4] [25] [35] [38] [39] [60].Source: §2 · pp. 2–3

MLLM as a judge

• Existing VLA evaluations emphasize task success rates, overlooking reasoning quality and beneficial trial-and-error; this work extends MLLM-as-a-Judge evaluation to VLAs.Prior MLLM-as-a-Judge work evaluates outputs with contextual prompts and can provide scores and rationales, including in multimodal settings [3] [62].Source: §2 · p. 3

3.1. Preliminaries

• Specialist VLAs learn actions from visual observations and instructions, while reasoning VLAs jointly predict reasoning and actions from those inputs.The reasoning-VLA training data also include multimodal samples represented with reasoning but no action target.Source: §3.1 · p. 3

3.2.1. Dual-Layers Data Pruning

• Dual-layer pruning retains reasoning at frames where scene-event and robot-motion cues coincide, removing repetitive low-entropy text that can dilute action learning.Scene boundaries come from a retrained event detector, while motion cues include abrupt velocity changes and gripper-state transitions; the parameter-free method retains reasoning only at jointly labeled keyframes.Source: §3.2.1 · p. 4

3.2.2. Dual-Teacher Adaptive Distillation

• Robot data receive action-teacher supervision, while multimodal reasoning data receive reasoning-teacher supervision, matching guidance to each modality.The specialist VLA provides fine-grained action supervision, while the reasoning-capable finetuning initialization preserves multimodal competence. Soft supervision reduces gradient conflicts; the auxiliary weight λ is 0.15.Source: §3.2.2 · pp. 4–5

3.3. VLA Score

• VLA Score evaluates reasoning, action smoothness, intention, and reason–act alignment alongside simulation outcome, using retrieved task and scene examples as context for a VLM judge.It remains compatible with models without explicit reasoning and can be adapted to different simulation environments.Source: §3.3 · pp. 5–6, Figure 3 · p. 5

4.1. Main Results

• On SimplerEnv, DUALVLA reaches 61.0 average success, +8.0 over InstructVLA-G, +5.0 over InstructVLA-E, and +3.9 over ThinkACT, with about 20% inference speedup.The passage attributes the improvements to resolving action degeneration and reports gains across multiple tasks and settings.Source: §4.1 · p. 6

4.2. Real-world Results

• In real-world tasks, DUALVLA raises average success from 45% to 60.0%, with gains in both Move Objects and Handover Objects.The evaluation used 10 rollouts per task; each task had 50 demonstration trajectories, and the robot produced 14-DoF dual-arm actions from three camera views.Source: §4.2 · p. 7, Table 3 · p. 7

4.3. VLA Score Results

• VLA Score assesses object recognition, action inference, reasoning-action consistency, and whether the robot moves toward the target.The passage frames this evaluation as an extension of video understanding.Source: §4.3 · p. 7
• DUALVLA achieves the highest VLA Score among reasoning VLAs, combining the reasoning teacher’s capability with smoother action generation that cross-entropy alone cannot achieve.The reported failure analysis finds that reasoning VLAs’ low action scores often reflect difficulty approaching or manipulating targets throughout a trajectory.Source: §4.3 · p. 7, Table 5 · p. 8

4.4. Ablation Study

• Teacher distillation improves action execution while preserving multimodal capability, and data distillation further helps by reducing excessive embodied reasoning; plain fine-tuning yields only marginal gains.The ablations compare the InstructVLA-G baseline and fine-tuning without pruning or distillation.Source: §4.4 · p. 8, Table 6 · p. 8
• Two-layer pruning outperforms single-layer pruning and random dropout; scene pruning contributes most of the gain, with action pruning adding a complementary improvement.The passage attributes the result to deliberate pruning rather than randomness.Source: §4.4 · p. 8, Table 7 · p. 8
• Dual-teacher adaptive distillation strengthens action generation while preserving multimodal understanding, whereas action-teacher-only supervision improves action slightly but substantially reduces multimodal capability.Reasoning-teacher distillation preserves multimodal reasoning without a significant difference from the reasoning teacher or base VLM; full single-teacher distillation offers no performance gain.Source: §4.4 · p. 8, Table 8 · p. 8, Figure 5 · p. 8
• OpenVLA-Dual surpasses OpenVLA in action performance and ECoT in multimodal generalization, though its action gains are modest given discrete-token action constraints.This extension tests the framework’s extrapolation to OpenVLA and ECoT.Source: Figure 6 · p. 8

• DUALVLA combines dual-layer pruning and dual-teacher adaptive distillation to mitigate action degeneration while strengthening action execution and preserving reasoning.Source: §5 · pp. 8–9
• The two-teacher dependency adds structural complexity, and distillation requires extra forward passes despite forward computation not being the main VLA training bottleneck.Both teachers are pretrained and add no extra training cost; preliminary attention-level distillation retained similar benefits with reduced computation.Source: §5 · p. 9
• Future work will simplify distillation, reduce reliance on teacher models, and extend DUALVLA to broader embodied settings.Source: §5 · p. 9

A. VLA Score

• The prompts are shown in Figures 15 and 16.Source: A. VLA Score · p. 13

Data Pruning

• Repetitive reasoning skews training toward low-information segments, suppressing the action objective; pruning these samples rebalances supervision toward diverse state–action pairs.The passage characterizes pruning as a way to adjust the effective training distribution, not merely reduce data.Source: B. Additional Qualitative Analyses · p. 13
• Robot reasoning samples form dense, low-entropy clusters, and consecutive changing frames can repeat the same intention, such as “Move Near.”Low-level movement steps often correspond to only a few distinct semantic intentions, while multimodal data remain more dispersed and semantically diverse.Source: Figure 7 · p. 13, B. Additional Qualitative Analyses · p. 13

Why does DUALVLA work?

• Naïvely mixing robot data favors reasoning-dominant updates and suboptimal action performance; pruning plus dual-teacher supervision steers optimization toward a Pareto-efficient balance.Pruning attenuates over-represented reasoning gradients, while expert-guided distillation aligns gradients without weakening action learning; either mechanism alone is insufficient.Source: B. Additional Qualitative Analyses · p. 13

C.1. Additional Visualization

• Figures 9 and 10 show successful DUALVLA execution examples on SimplerEnv.Source: §C.1 · p. 13
• Figures 11 and 12 show successful DUALVLA execution examples on real-world tasks.Source: §C.1 · p. 13
• Consecutive robot frames can repeat identical reasoning such as “Move Near” despite scene and motion changes, motivating pruning.Source: Figure 8 · p. 14

C.2. Additional Failure Case Analysis

• DUALVLA failures primarily reflect insufficient action detail causing imprecise or unstable execution and missing historical context leading to suboptimal or inconsistent decisions.Some failures also reflect simulation bias: tasks that are not perfectly completed may still be marked unsuccessful, underscoring the importance of fine-grained assessment through VLA Score.Source: §C.2 · p. 14