DialogVLA: Bridging Reasoning Action in Bimanual Manipulation via Parallel Mixture-of-Experts

DialogVLA Overall We present DialogVLA, a dialog reasoning-based vision-language-action model for bimanual robot coordination. The model generates structured dialog reasoning to decompose tasks into per-arm subtasks, routes through Parallel MoE to select skill and arm experts, and produces coordinated bimanual actions via Flow Matching. DialogVLA achieves state-of-the-art performance across RoboTwin simulation, RoboCasa simulation, and real-world bimanual manipulation tasks

55.3%

RoboCasa GR-1

+6.5%p over Qwenπ₀

82.8%

RoboTwin Easy

+32.3%p over π₀

63.9%

RoboTwin Hard

+46.7%p over π₀

75.4%

Real-world success

8 bimanual tasks

Abstract

Structurally coupling reasoning and action

Vision-Language-Action (VLA) models show strong generalization for instruction-based manipulation, yet in bimanual settings two challenges remain: misalignment between high-level decision-making and low-level action generation, and uncertainty in arm-role assignment. Existing methods implicitly project manipulation stages and arm-leadership decisions into a single action space, which fails to secure structural control stability in complex collaborative situations.

We propose DialogVLA, which explicitly reasons about manipulation stages and arm roles through Dialog Reasoning, and binds high-level reasoning to continuous control by conditioning it on MoE-based structural routing and a Flow Matching Action Expert. A Two-Stage training strategy separates reasoning/routing from action generation to improve stability and generalization.

On RoboCasa and RoboTwin benchmarks and in real-robot experiments, DialogVLA consistently outperforms prior robot-learning and VLA baselines — reaching 82.8% / 63.9% average success on RoboTwin Easy / Hard, and 75.4% on real-world bimanual collaborative tasks.

Contributions

What's new

A structural VLA framework that explicitly decomposes Skill Phase and Arm Role for bimanual manipulation.

A design that models collaborative states by configuring Skill MoE and Arm MoE in parallel, delivering them directly as conditioning signals to action generation.

A Two-Stage Training strategy that separates reasoning/routing learning from action-generation learning, validated in both simulation and real-robot environments.

Positioning

Why not just another VLA?

Three ways to get from reasoning to action. The first hides the decision; the second exposes it but breaks the gradient. DialogVLA keeps one model and carries the decision forward as a continuous signal.

(a) Standard VLA

One implicit action space

Maps observation and instruction straight to actions
Skill phase and arm role are never named
Errors accumulate over long-horizon bimanual tasks

(b) Hierarchical VLA

Two models, one text bottleneck

A planner emits a discrete language subtask
A separate low-level policy consumes that string
Everything not expressible in words is lost at the seam

One model, continuous coupling

Skill Phase and Arm Role stay explicit and separable
Routing probabilities condition action generation directly
No discretization step between deciding and acting

Method

How DialogVLA works

DialogVLA architecture — **DialogVLA Architecture.** Multi-view images (head, left/right wrist) and a language instruction are encoded by the VLM. The **Dialog Mixture-of-Experts** module produces dialog reasoning tokens \(h_{\text{Dialog}}\) and MoE conditioning vectors \(h_{\text{MoE}}\) through \(N\) transformer blocks with parallel Skill and Arm routers. The **Action Expert** takes the conditioning vector \(h_{\text{cond}}\), proprioceptive state \(\mathbf{p}\), and a noisy action chunk, then denoises it into coordinated bimanual actions via shared attention with the VLM.

Step 1

Dialog Reasoning

From the head-view image and the instruction, the VLM autoregressively emits a structured dialog that names the current Skill Phase and assigns a role to each arm — making the decision explicit instead of burying it in the action vector.

Step 2

Parallel MoE routing

Every FFN in the backbone becomes two expert groups. A Skill Router picks the top-2 of 7 phase experts; an Arm Router picks 1 of 3 role experts. Their probabilities form a 10-dim conditioning vector.

Step 3

Conditioned action

The projected routing vector modulates the Flow Matching Action Expert through AdaRMS, so the velocity field it generates matches the current phase and arm role — even under identical visual input.

Dialog Mixture-of-Experts routing — **Dialog Mixture-of-Experts.** At each timestep, Dialog Reasoning emits a structured JSON that decomposes the task into a subtask with per-arm think/act traces, a global plan, and a coordination note. The resulting hidden state is fed to two parallel routers: the **Arm Router** activates a single expert (\(K{=}1\)) among Left, Bimanual, and Right, while the **Skill Router** activates the top-two experts (\(K{=}2\)) among a shared expert and skill-specific experts such as Lift, Grasp, and Place. The 3-dim arm and 7-dim skill probabilities are concatenated into a 10-dim MoE conditioning vector, which the Conditioning Projector expands to 1024 dimensions.

Two-stage training pipeline — **Two-Stage training pipline of DialogVLA.** Stage 1 trains the Dialog MoE (Skill/Arm routers and LM Head) for dialog reasoning and MoE conditioning, with the Conditioning Projector and Action Expert frozen. Stage 2 freezes the VLM and Dialog MoE, training only the Conditioning Projector and Action Expert to produce bimanual actions from the learned routing signals.

Rollouts

RoboCasa GR-1 Tabletop

Bottle to cabinet close

Can to drawer close

Cuttingboard to basket

Cuttingboard to pan

Placemat to basket

Placemat to bowl

Tray to cardboard box

Tray to plate

Results

RoboCasa GR1 Tabletop

**Table.** RoboCasa GR1 Tabletop (verb-grouped averages, success rate %), 50 rollouts per task on 24 PnP tasks.
Task (Avg)	Isaac-GR00T	QwenGR00T	QwenPI	QwenOFT	QwenFAST	VisionOnly	DialogVLA (Ours)
PnP * to * Close (Avg)	24.2	50.3	42.3	43.7	35.0	54.0	44.6
PnP Novel From Cuttingboard To * (Avg)	56.9	52.8	46.0	50.4	50.4	45.2	57.6
PnP Novel From Placemat To * (Avg)	51.9	38.0	43.5	41.5	33.5	32.5	53.5
PnP Novel From Tray To * (Avg)	55.1	39.2	44.0	49.2	32.0	44.8	64.0
PnP Novel From Plate To * (Avg)	57.6	58.5	44.0	61.0	45.0	42.0	56.8
Overall Average	47.6	47.8	43.9	48.8	39.0	44.7	55.3

Rollouts

RoboTwin 2.0

Easy Scenario

Dump bin bigbin

Handover block

Grab roller

Pick dual bottles

Hard Scenario

Dump bin bigbin

Handover block

Grab roller

Pick dual bottles

Results

RoboTwin 2.0 benchmark

**Table.** RoboTwin 2.0 results on 12 bimanual tasks (success rate %). DialogVLA achieves the highest performance on all 12 tasks.
Task	ACT		π₀		RDT		DialogVLA (Ours)
Task	Easy	Hard	Easy	Hard	Easy	Hard	Easy	Hard
Dump Bin Bigbin	68.0	1.0	83.0	24.0	64.0	32.0	94.0	55.0
Grab Roller	94.0	25.0	96.0	80.0	74.0	43.0	98.0	83.0
Handover Block	42.0	0.0	45.0	8.0	45.0	14.0	88.0	67.0
Lift Pot	88.0	0.0	84.0	36.0	72.0	9.0	92.0	59.0
Pick Dual Bottles	31.0	0.0	57.0	12.0	42.0	13.0	82.0	63.0
Place Bread Skillet	7.0	0.0	23.0	1.0	5.0	1.0	91.0	77.0
Place Cans Plasticbox	16.0	0.0	34.0	2.0	6.0	5.0	57.0	18.0
Place Dual Shoes	9.0	0.0	15.0	0.0	4.0	4.0	77.0	63.0
Put Object Cabinet	15.0	0.0	68.0	18.0	33.0	18.0	66.0	67.0
Scan Object	2.0	0.0	18.0	1.0	4.0	1.0	72.0	54.0
Stack Blocks Three	0.0	0.0	17.0	0.0	2.0	0.0	86.0	82.0
Stack Bowls Three	48.0	0.0	66.0	24.0	51.0	17.0	91.0	79.0
Overall Average	34.5	2.2	50.5	17.2	33.5	13.1	82.8	63.9

Each task is evaluated with 100 rollouts on the 14-DoF ALOHA-AgileX embodiment.

Real Robot

Real-world bimanual tasks

Evaluated on the SO-ARM 101 platform (head + left/right wrist cameras) across 8 bimanual collaborative tasks. DialogVLA reaches 75.4% average success, outperforming ACT, Diffusion Policy, SmolVLA, π₀, and GR00T-1.5N on every task.

Handover banana