Neuromorphic Control of a Flapping-Wing Robot on Resource-Constrained Hardware
Rim El Filali, Chenrui Feng, Chao Gao, Weibin Gu
Abstract
Flapping-Wing Micro Aerial Vehicles (FWMAVs) provide exceptional maneuverability and aerodynamic efficiency but pose significant challenges for onboard control due to nonlinear dynamics and stringent Size, Weight, and Power (SWaP) constraints, as exemplified by a butterfly-inspired robot less than 30 gram. To this end, we present a hierarchical neuromorphic control framework that enables fully onboard, closed-loop flight on a widely available, resource-constrained ESP32 microcontroller with a unit cost of approximately $5. Specifically, our method deploys two lightweight Spiking Neural Networks (SNNs) onboard: one for state estimation from raw sensory feedback and another for control via modulation of a Central Pattern Generator (CPG) for wing actuation. Trained by imitation learning, the system achieves stable pitch and heading angle tracking during untethered real-world flight. Experimental results further reveal that the SNN-based controller reduces latency by 36% (1059us to 680us) and power by 18% (0.033W to 0.027W) for inference compared to the conventional Artificial Neural Network (ANN) baseline, demonstrating the viability of spike-based computation without specialized hardware. To the best of our knowledge, this work constitutes the first demonstration of fully onboard neuromorphic control for autonomous flight of a FWMAV, highlighting the potential of SNNs to enable energy-efficient autonomy under stringent SWaP constraints. Visual abstract: http://bit.ly/4nI8ECY
AI Impact Assessments
(1 models)Scientific Impact Assessment
1. Core Contribution
This paper presents what the authors claim is the first fully onboard neuromorphic flight controller for a sub-30 gram butterfly-inspired flapping-wing micro aerial vehicle (FWMAV). The system deploys two lightweight Spiking Neural Networks (SNNs) on a ~$5 ESP32-S3 microcontroller: one for state estimation from raw IMU data and another for attitude control via modulation of a Central Pattern Generator (CPG). The key novelty lies in the intersection of three elements: (1) neuromorphic computing applied to a particularly challenging control domain (underactuated, nonlinear flapping-wing flight), (2) deployment on extremely resource-constrained commodity hardware rather than specialized neuromorphic chips, and (3) real-world untethered flight validation.
The hierarchical architecture — separating rhythmic wing actuation (CPG) from corrective control (SNN) — is a well-motivated design choice. The paper convincingly demonstrates that the CPG-aware controller generalizes better than the CPG-agnostic variant, as the latter must learn both rhythmic patterns and corrections simultaneously.
2. Methodological Rigor
The methodology is generally sound but has notable limitations. The training approach uses behavioral cloning (imitation learning) from a Madgwick estimator + PID controller, meaning the SNN can only match — not exceed — the baseline's performance. The evaluation metrics confirm this: RMSE values are computed against the Madgwick/PID baselines rather than ground truth, making it difficult to assess absolute performance quality.
The computational benchmarks (36% latency reduction, 18% power reduction) are the most concrete results, though the absolute numbers are modest in context — the power savings amount to 6 mW, and the latency improvement is ~379 μs. Importantly, a "conventionally implemented" SNN (without event-driven optimizations) performs comparably to the ANN, suggesting the gains come primarily from the sparse event-driven implementation rather than inherent SNN properties.
The experimental validation includes real untethered flight, which is commendable, but the paper lacks quantitative flight performance comparisons (e.g., SNN-controlled flights vs. PID-controlled flights in terms of tracking accuracy, flight duration, or stability margins). The free-flight results (Table 1) show RMSE values but only against the baseline controller's outputs from the same logs, not head-to-head comparisons under identical conditions.
Data augmentation with ~150 minutes of synthetic data to supplement ~10 minutes of real flight data is practical but raises questions about domain gap, which are not thoroughly addressed.
3. Potential Impact
The work has several impact vectors:
Practical accessibility: The open-sourced *Espikify* tool for converting PyTorch SNNs to ESP32-ready C code lowers the barrier for deploying neuromorphic systems on commodity hardware. This could catalyze experimentation in resource-constrained robotics beyond FWMAVs.
Bio-inspired robotics: Demonstrating that SNNs can replace conventional estimation and control on an extremely weight-constrained platform validates a path toward biologically plausible control architectures for micro-robots.
Neuromorphic computing: This serves as a real-world proof-of-concept that SNNs can provide tangible efficiency benefits on non-neuromorphic hardware, though the gains are incremental rather than transformative on conventional processors.
However, the impact is somewhat constrained by the fact that the SNN merely mimics a PID controller. The system doesn't demonstrate capabilities beyond what conventional control achieves — it achieves comparable performance with modest efficiency gains. The deeper question of whether neuromorphic control can enable *qualitatively different* capabilities remains unanswered.
4. Timeliness & Relevance
The paper is well-timed. Interest in both FWMAVs and neuromorphic computing is growing, and the intersection remains underexplored — particularly for real hardware demonstrations. The focus on commodity hardware (ESP32) rather than specialized neuromorphic chips (e.g., Intel Loihi, BrainChip Akida) is pragmatically motivated, as these chips are either expensive, power-hungry at the system level, or unavailable at the required form factor for micro-robots.
The SWaP constraints of sub-30g flight platforms represent a genuine bottleneck where every milliwatt matters, making this a relevant testbed for efficient computing approaches.
5. Strengths & Limitations
Strengths:
Limitations:
Additional observations:
The paper would benefit from a clearer articulation of *when* neuromorphic control becomes essential rather than merely convenient. On the ESP32, the demonstrated SNN runs at 680 μs per inference — both this and the ANN's 1059 μs are well within the 10 ms (100 Hz) control loop budget. The efficiency gains, while real, don't unlock new capabilities in this particular application.
The comparison with [19] (Stroobants et al.) on quadrotors is helpful for positioning but also highlights that this paper's contribution is more about platform novelty (FWMAV + ESP32) than methodological advancement. The SNN architectures are relatively standard CUBA-LIF networks trained with surrogate gradients.
Summary
This is a solid systems-level integration paper that achieves a meaningful "first" in deploying neuromorphic control for untethered FWMAV flight on commodity hardware. The engineering contribution is clear, the open-source tools add practical value, and the real-world demonstration is credible. However, the scientific depth is limited by the imitation-only training paradigm, modest efficiency margins, and lack of ground-truth validation or demonstration of capabilities beyond existing conventional approaches.
Generated May 20, 2026
Comparison History (21)
Paper 2 likely has higher impact due to broad relevance: scalable humanoid skill learning from monocular video addresses a central, timely robotics problem with wide applicability (humanoids, animation, VR/AR, biomechanics). The proposed Direct Dynamic Retargeting removes a known geometric-bias bottleneck and integrates task-space optimization with contact-rich MPC, offering methodological innovation and a general pipeline that can feed downstream RL. Paper 1 is novel and strong in real-world embedded neuromorphic flight, but its impact is narrower to FWMAV/SNN-on-MCU niches compared to the cross-domain significance of video-to-humanoid dynamics.
Paper 1 addresses a critical safety bottleneck in the rapidly expanding field of foundation models for embodied AI. By establishing a quantitative link between reasoning consistency and physical trajectory safety, it offers a widely applicable framework for VLA runtime monitoring. While Paper 2 presents an impressive first-of-its-kind engineering milestone in neuromorphic micro-robotics, Paper 1's insights into VLM fragility and interpretability will likely drive broader, cross-disciplinary impact across AI safety, autonomous driving, and general robotics.
Paper 2 demonstrates a first-of-its-kind achievement: fully onboard neuromorphic control for autonomous FWMAV flight using commodity hardware. This combines novelty (SNNs on resource-constrained platforms), real-world hardware validation, and cross-disciplinary impact spanning neuromorphic computing, robotics, and bio-inspired flight. The practical implications for micro-robotics and edge AI are substantial. Paper 1, while technically solid, represents an incremental advance in crowd navigation by combining existing concepts (CVaR, barrier functions, RL). Paper 2's pioneering demonstration and broader applicability across fields give it higher potential impact.
Paper 1 is likely to have higher impact because it introduces a low-cost, reproducible real-world benchmark for VLA models—an enabling infrastructure contribution that can standardize evaluation across many labs, accelerate progress, and influence a broad robotics/ML community. Its applicability spans manipulation, benchmarking methodology, domain adaptation, and reproducibility initiatives, making it timely amid rapid VLA development. Paper 2 is novel and well-motivated, but its impact is more specialized (FWMAV + SNN control) and may affect a narrower set of researchers compared to a widely adoptable benchmark.
Paper 1 demonstrates a first-of-its-kind neuromorphic control system for flapping-wing MAVs on commodity hardware, combining novelty (SNNs on $5 microcontroller), real-world experimental validation, and quantified improvements in latency and power. It bridges neuromorphic computing, bio-inspired robotics, and embedded systems with immediate practical implications for SWaP-constrained autonomous platforms. Paper 2 contributes a useful data generation pipeline but is more incremental, combining existing techniques (LLMs, 3DGS, trajectory planning) without novel algorithmic contributions, and lacks real-world validation of downstream task performance.
Paper 1 demonstrates the first fully onboard neuromorphic control for autonomous FWMAV flight, combining novel SNN-based control with practical deployment on a $5 microcontroller. It addresses critical SWaP constraints with measurable improvements (36% latency reduction, 18% power savings) and has broad implications for neuromorphic computing, micro-robotics, and edge AI. Paper 2 makes an incremental contribution to a niche SLAM subfield (ground texture mapping in low-dynamic environments), with narrower applicability and less transformative potential.
Paper 2 likely has higher scientific impact due to broader relevance and timeliness: it targets vision-language-action modeling, a fast-moving area with wide applicability in robotics, embodied AI, and multimodal ML. Its systematic, unified comparison of latent action supervision strategies offers generalizable insights (formulation-task correspondence) and provides code, improving reproducibility and adoption. Paper 1 is novel and rigorous with strong real-world demonstration under SWaP constraints, but its impact is narrower (FWMAV control and SNN deployment on a specific microcontroller), limiting cross-field reach compared to VLA training methodology.
Paper 2 likely has higher impact due to its strong novelty and real-world demonstration: the first fully onboard neuromorphic (SNN) closed-loop control of an untethered flapping-wing MAV on a $5, resource-constrained MCU, with quantified latency/power gains and no specialized neuromorphic hardware. This directly advances deployable autonomy under SWaP constraints and bridges robotics, neuromorphic computing, embedded systems, and control. Paper 1 is timely and useful for smoother RL control, but is more incremental within RL algorithm design and appears primarily benchmark-driven, with less clear hardware-validated deployment significance.
Paper 1 demonstrates the first fully onboard neuromorphic control for autonomous FWMAV flight using commodity hardware ($5 ESP32), achieving significant latency (36%) and power (18%) reductions. This represents a groundbreaking intersection of neuromorphic computing, bio-inspired robotics, and edge AI with broad implications for SWaP-constrained systems. Paper 2, while solid, offers an incremental improvement in semantic navigation via heatmap representations. Paper 1's novelty—deploying SNNs on non-specialized hardware for real-world flight—opens new research directions across multiple fields including neuromorphic engineering, micro-robotics, and embedded AI.
Paper 1 demonstrates a first-of-its-kind achievement: fully onboard neuromorphic control for autonomous FWMAV flight using commodity hardware (~$5 ESP32). This combines novelty (SNN-based control on resource-constrained platforms), practical impact (significant latency and power reductions), and cross-disciplinary relevance (robotics, neuromorphic computing, bio-inspired design). Paper 2 presents a solid methodological contribution to information-theoretic exploration, but its advances are more incremental within an established RL/experimental design framework. Paper 1's tangible hardware demonstration and accessibility make it more likely to inspire follow-up work across multiple communities.
Paper 1 demonstrates a first-of-its-kind achievement: fully onboard neuromorphic control for autonomous FWMAV flight on a $5 microcontroller. It bridges neuromorphic computing, bio-inspired robotics, and edge AI with concrete hardware results showing 36% latency and 18% power reductions. This novelty across multiple fields (neuromorphic engineering, micro-robotics, embedded systems) and the practical demonstration on extreme SWaP-constrained platforms gives it broader impact potential. Paper 2, while solid, represents a more incremental combination of known techniques (GNNs + NMPC) for multi-robot planning.
Paper 2 has higher impact potential: it reports a first fully onboard neuromorphic (SNN) closed-loop control demonstration for an untethered flapping-wing robot on a commodity $5 ESP32, directly addressing SWaP constraints with real-flight validation and measurable latency/power reductions. This is timely (edge AI/neuromorphic autonomy) and broadly relevant to robotics, embedded systems, and neuromorphic computing. Paper 1 is methodologically strong and useful for differentiable tactile simulation, but its impact is more niche (tactile rendering/simulation pipelines) and relies on paired high/low-res simulation data rather than a compelling real-world deployment milestone.
Paper 1 presents a significant breakthrough in real-world robotics by deploying energy-efficient Spiking Neural Networks for the control of highly constrained flapping-wing micro aerial vehicles. Its demonstration of reduced latency and power consumption on low-cost hardware has strong implications for physical robotics and edge AI. In contrast, Paper 2 offers a simulation benchmark for game-inspired morphologies, which, while useful for game AI and animation, has more limited real-world scientific and engineering impact.
Paper 2 likely has higher scientific impact due to its strong novelty and real-world significance: it demonstrates (apparently first) fully onboard neuromorphic (SNN) closed-loop control for an untethered flapping-wing robot on a $5, resource-constrained microcontroller, with measured latency and power gains. This directly advances deployable autonomy under SWaP constraints and is timely for edge robotics/neuromorphic computing, with cross-field relevance (aerial robotics, embedded ML, neuromorphic systems). Paper 1 is impactful within robot learning, but is a more incremental algorithmic refinement in an already active RL adaptation landscape.
Paper 2 presents a pioneering demonstration of fully onboard neuromorphic control for flapping-wing micro aerial vehicles under extreme size, weight, and power constraints. By achieving real-world untethered flight using Spiking Neural Networks on a cheap, resource-constrained microcontroller, it demonstrates significant novelty, broad cross-disciplinary impact in neuromorphic engineering and robotics, and high potential for real-world application in miniature autonomous systems. Paper 1 offers a useful but more incremental heuristic approach for RL reward tuning.
Paper 1 demonstrates a first-of-its-kind fully onboard neuromorphic control system for flapping-wing MAVs on commodity hardware ($5 ESP32), achieving meaningful latency and power reductions. This bridges neuromorphic computing, bio-inspired robotics, and embedded systems with clear real-world applicability and opens new research directions for SNN deployment without specialized neuromorphic chips. Paper 2 contributes a valuable evaluation framework for manipulation, but is primarily methodological/diagnostic rather than enabling new capabilities. Paper 1's novelty as a first demonstration with cross-disciplinary impact gives it higher potential scientific influence.
Paper 1 presents a highly novel breakthrough as the first demonstration of fully onboard neuromorphic control for a flapping-wing micro aerial vehicle on ultra-low-cost, resource-constrained hardware. Its integration of Spiking Neural Networks for both state estimation and control offers significant implications for energy-efficient edge AI and bio-inspired robotics. Paper 2, while methodologically sound and practically useful for multi-robot systems, represents a more incremental advancement (warm-starting an optimizer with LSTMs) compared to the pioneering system-level innovation in Paper 1.
Paper 1 is more novel and broadly impactful: it demonstrates fully onboard neuromorphic (SNN) state estimation and control for real flapping-wing flight on a commodity $5 ESP32, with measured latency/power reductions versus ANN and real-world validation under extreme SWaP constraints. This advances practical neuromorphic autonomy and can influence robotics, embedded ML, and neuromorphic computing communities. Paper 2 is timely and application-relevant but is primarily a simulated planner with incremental architectural additions (attention + safety module) and narrower cross-field novelty; lack of real-world deployment reduces immediate impact.
Paper 2 likely has higher impact due to broader relevance and reusability: it introduces a system-level, real-world benchmark combining dexterous manipulation, agentic perception, and closed-loop decision routing, with a dataset and standardized evaluation that can be adopted across robotics and multimodal/agent research. This can catalyze comparable progress and cross-lab benchmarking. Paper 1 is novel and rigorous with strong real-world constraints and a first demonstration claim, but its impact is more specialized to FWMAV control and neuromorphic embedded inference. Benchmarking infrastructure tends to propagate faster and wider.
Paper 2 presents a pioneering, first-of-its-kind demonstration of fully onboard neuromorphic control for a flapping-wing micro aerial vehicle on non-specialized hardware. Its integration of Spiking Neural Networks to solve severe Size, Weight, and Power (SWaP) constraints offers a significant breakthrough in bio-inspired robotics and energy-efficient autonomy. While Paper 1 provides a valuable fine-tuning method for navigation foundation models using established techniques, Paper 2's demonstrable reductions in latency and power for a highly complex, nonlinear control problem suggest broader, disruptive implications for micro-robotics and neuromorphic computing.