Expertise need not monopolize: Action-Specialized Mixture of Experts for Vision-Language-Action Learning

• Conventional coupled routing makes load balancing and task-specific expert specialization compete, so strong balancing can hurt performance while weak balancing risks expert collapse.Source: §1 · p. 2
• AdaMoE uses an independent scale adapter alongside the traditional router, selecting experts by relevance while controlling their contribution weights separately.Source: §1 · p. 2
• 1.8% improvement over π0 on LIBERO, 9.3% success-rate gain on 19 RoboTwin hard-setting tasks, and 21.5% improvement in real-world experiments are reported.Source: §1 · p. 3
• The approach inherits pretrained VLA foundation-model weights to extend models into MoE at low cost, a valuable strategy amid scarce robot data.Source: §1 · pp. 2–3

2.1 PROBLEM FORMULATION

• The policy takes multi-view images, a language instruction, and robot state as observations and predicts a chunk of future actions using conditional flow matching.The formulation builds on the pretrained flow-matching VLA model π0.Source: §2.1 · pp. 3–4

2.2 MOE-ARCHITECTURE

• The action expert uses shared experts for common action patterns and routed experts for task-specific specialization; shared experts inherit original FFN weights and routed experts are copies.Shared experts are always active, while top-k routing activates selected routed experts for each token.Source: Figure 1 · p. 3, §2.2 · p. 4
• Load balancing encourages uniform routed-expert use to prevent collapse, and its loss is combined with the original flow-matching loss in training.The balancing term also encourages selection-frequency and gating-probability balance across routed experts.Source: §2.2 · pp. 4–5

2.3 DECOUPLED EXPERT SELECTION AND WEIGHTING

• In conventional MoE, load balancing and task-specific specialization compete through the same router, constraining expert specialization for manipulation tasks.The router's selection probabilities are also used as output weights.Source: §2.3 · p. 5
• AdaMoE separates top-k expert selection from contribution weighting, using a router to select experts and a scale adapter to adjust their output weights.The separated controls let the model select an expert for relevance while controlling its contribution independently.Source: §2.3 · p. 5

3.1 SIMULATION BENCHMARKS

• The simulation evaluation covers four LIBERO suites and 19 RoboTwin 2.0 tasks, each with 100 clean and 400 domain-randomized expert trajectories.Source: §3.1 · p. 6

3.2.1 Q1: DOES MOE IMPROVE UPON DENSE VLA MODELS?

• 1.8% on LIBERO (94.2% → 96.0%) and 9.3% on RoboTwin (40.4% → 49.7%) exceed baseline results across four suites and 19 tasks.The improvements are more pronounced on the large-scale RoboTwin benchmark.Source: Table 1 · p. 6, Table 2 · p. 6, §3.2.1 · p. 6

3.2.2 Q2: DO OUR MOE EXPERTS ACHIEVE MEANINGFUL SPECIALIZATION?

• Expert 3’s token use rises during target positioning and gripper release, and similar activation patterns across tasks align with manipulation phases.The analysis measures each expert’s share of tokens at each frame.Source: §3.2.2 · pp. 6–7, Figure 2 · p. 7

3.2.3 Q3: HOW EFFECTIVE IS OUR DECOUPLED ARCHITECTURE DESIGN?

• 1.6% average improvement over vanilla MoE: additive AdaMoE leads across LIBERO suites, while concatenated CSMoE shows moderate improvements.The ablation compares vanilla MoE, concatenated scale adaptation, and additive scale-adapter weights.Source: §3.2.3 · p. 7, Table 3 · p. 8

3.3 HYPER-PARAMETER ABLATION STUDIES

• 96.0% is reported for k=1, four experts, and λbalance=0.01; inadequate and excessive balancing reach 94.5% and 95.1%, respectively.k=2 reaches 95.4%; four experts outperform eight (95.6%), and inadequate balancing lowers LIBERO-Long from 92.0% to 88.0%.Source: §3.3 · p. 8

3.4.1 EXPERIMENTAL SETUP

• The real-world setup uses a dual-arm Piper platform for four tasks, initializes from RoboTwin, post-finetunes on 450 demonstrations, and evaluates each task over 50 trials against π0.Source: §3.4.1 · pp. 8–9

3.4.2 RESULTS

• AdaMoE improves over π0 on all four real-world tasks, with particularly notable gains in complex scenarios requiring precise coordination.The reported advantages persist under sensor noise, lighting variations, and object-pose uncertainty.Source: Table 5 · p. 9, §3.4.2 · p. 9

4.1 VISION–LANGUAGE–ACTION MODELS FOR ROBOT MANIPULATION

• Flow-matching VLAs enable parallel action-chunk decoding with lower latency, but scaling their action experts while preserving strict latency requirements remains underexplored.The section contrasts this with autoregressive decoding, whose inference latency grows roughly linearly with action horizon.Source: §4.1 · pp. 9–10
• VLA models leverage pretrained vision-language backbones for instruction following and visual grounding, but remain modest in size under real-time control constraints.Autoregressive decoding offers scaling precedent but grows linearly in latency with action horizon; flow-matching action-expert scaling is comparatively underexplored.Source: §4.1 · p. 9

4.2 MIXTURE-OF-EXPERTS ARCHITECTURES IN DEEP LEARNING

• Existing MoE methods for VLAs couple expert selection with contribution weights and lack efficient pathways to scale well-pretrained VLA models.Sparse expert activation can scale model capacity while maintaining computational efficiency.Source: §4.2 · p. 10
• Sparse MoE activates only selected expert modules, and prior work includes dense-checkpoint upcycling and robotics applications to multitask learning and trajectory prediction.The section notes Sparse Upcycling's reported use of 50% of original pretraining cost, alongside MENTOR, Tra-MoE, and VER.Source: §4.2 · p. 10

5 CONCLUSION

• 1.8% on LIBERO tasks, 9.3% on RoboTwin 2.0 domain-randomized tasks, and 21.5% across four real-world manipulation tasks are reported improvements over the π0 baseline.The conclusion presents these as comprehensive evaluation gains.Source: §5 · p. 10
• AdaMoE uses an independent scale adapter alongside the router to select experts by relevance while independently controlling their contribution weights.The design combines decoupled selection and weighting with weight inheritance from pretrained dense VLA models.Source: §5 · p. 10
• The authors present AdaMoE as a practical pathway for scaling pretrained VLA models while addressing the trade-off between load balancing and performance.They frame collaborative expert utilization as matching the multifaceted nature of robotic manipulation.Source: §5 · p. 10

A.1 USAGE OF LARGE LANGUAGE MODELS

• Claude (Anthropic) was used as a writing assistant to improve the manuscript's language quality and readability.The stated purpose was refining sentence structure, enhancing clarity, and polishing academic writing style.Source: §A.1 · p. 15
• The authors state that Claude was employed solely for sentence-structure refinement, clarity, and academic writing polish.Source: §A.1 · p. 15
• The authors state that all technical content, experimental results, and scientific contributions are their original work.Source: §A.1 · p. 15

A.2 DETAILS OF REAL-WORLD EXPERIMENTS

• Real-world validation covered four manipulation tasks on an AgileX Cobot Magic mobile platform in an ALOHA configuration.The platform has four robotic arms; each arm is a six-degree-of-freedom AgileX Piper with a one-degree-of-freedom parallel gripper.Source: §A.2 · p. 15
• The platform uses a 640 × 480 RGB-D camera at approximately 30 Hz plus two wrist-mounted RGB-D cameras for additional end-effector views.Source: §A.2 · p. 15
• The stack-plate task varies bowl positions and tests two color-position arrangements equally often to assess robustness across spatial configurations.The task requires coordinating both arms to stack colored bowls in a designated sequence.Source: §A.2 · pp. 15–16
• The bell task requires precise activation despite nearly identical initial and final visual observations, so the experiment relaxed time limits and enforced strict spatial accuracy.The principal visual difference between successful-trajectory endpoints is the transient bell activation.Source: §A.2 · p. 16
• The cup task tests transparent-object handling and long-horizon bimanual coordination through a sequential two-arm handoff across the platform.Source: §A.2 · p. 16

• AdaMoE is trained for 120,000 steps.Source: §A.3 · p. 16
• Training uses the AdamW optimizer.Source: §A.3 · p. 16
• The router and base components use differentiated learning rates.Source: §A.3 · p. 16, Table 6 · p. 16

• Expert usage intensity is the proportion of tokens assigned to an expert at a frame.Source: §A.4 · p. 17
• The token count for expert i is measured at denoising step s for frame t against the total tokens at that frame.Source: §A.4 · p. 17
• The flow-matching inference process uses 10 equally spaced denoising steps.Source: §A.4 · p. 17

• RoboTwin 2.0 tasks are specified with a manipulation objective, a schema for object and end-effector placeholders, and paraphrased instruction examples.Source: §A.5 · p. 17, Table 7 · p. 17
• The language specification and cluttered tabletop scenarios are described as supporting generalizable manipulation across linguistic expressions and environmental conditions.Source: §A.5 · p. 17
• Figure 5 depicts bowl stacking, bottle adjustment, bell pressing, and cup placement demonstrations.Source: Figure 5 · p. 18
• Figure 6 shows RoboTwin 2.0 demonstrations including beat block hammer, placing dual shoes, handing over a block, and stacking two blocks.Source: §A.5 · p. 17, Figure 6 · p. 19
• Figure 7 presents LIBERO manipulation task demonstrations.Source: §A.5 · p. 17, Figure 7 · p. 20