{"id":"476ef244-5daa-47fc-96db-8e3d0b24eb62","arxiv_id":"2411.13945","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A spiking neural network trained by imitation can perform attitude estimation and control of a tiny quadrotor on real hardware, tracking commands within 3.03 degrees on average versus 2.67 degrees for the standard PID stack.","lead":"Researchers trained a brain-inspired, pulse-based neural network to guide a small drone's orientation using only raw sensor readings, and flew it onboard at 500 times per second. It tracked commanded angles almost as accurately as the standard controller, a step toward ultra-low-power neuromorphic autopilots.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 3.03° tracking figure is computed against the SNN's own internal attitude estimate, with no motion-capture ground truth on attitude, so the quantitative 'estimation and control' comparison is not yet supported.","rationale":"I read the central claim as a feasibility demonstration: a trained spiking neural network can replace the complementary-filter estimator and PID controller on a real quadrotor, mapping raw IMU measurements to motor commands at 500 Hz. The released code and data, the real flight tests, and the stable position response in Figure 4 are genuine evidence that the closed-loop control mapping works. The soft spot is the estimation component and its quantitative evaluation. The metric in Table I is defined as tracking of the estimated roll angle; for the SNN that estimate is an internal, unsupervised intermediate representation, never compared to a measured orientation. Without a ground-truth attitude measurement, the 3.03° figure cannot be interpreted as physical attitude-tracking error, and the title-level claim of attitude estimation is not validated. This is not a charge of dishonesty; the authors disclose the OptiTrack limitation, but the conclusion drawn from the metric is stronger than the evidence supports. This concern does not overturn the core feasibility result, so the conditional verdict remains appropriate, with the explicit condition being an external attitude validation or a more carefully scoped claim.","tokens_in":13119,"tokens_out":3783,"duration_ms":40743,"concrete_test":"Log full 6-DoF OptiTrack pose during the ten roll step-response flights and recompute Table I using motion-capture roll as ground truth for both the SNN and the PID stack. Also compute the RMSE between the SNN's internal roll estimate and the motion-capture roll. If the internal estimate deviates from true roll substantially more than the PID's complementary-filter estimate does, or if the true-roll RMSE for the SNN is much larger than 3.03°, the estimation claim and the quantitative comparison fail as stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-B defines the headline metric as the RMSE between the commanded roll setpoint and the 'resulting (estimated) roll angle,' and Table I uses this to report 3.03° for the SNN versus 2.67° for the PID stack. For the SNN, that estimated roll angle is an internal, unsupervised intermediate representation: Section II-C.3 explains that the estimation sub-network is trained only through the merged control loss, and Section II-D states that OptiTrack provides only position and absolute heading, not full attitude. The network may therefore have learned a control-relevant latent variable that is not a physical attitude estimate, and its self-scored error can be small even if the actual quadrotor attitude differs substantially. Since the title and contribution claim both 'attitude estimation and control,' and the quantitative comparison to the PID stack rests on this unvalidated internal quantity, the central claim is overreaching until the estimate is checked against true attitude or the claim is narrowed to end-to-end control.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a spiking neural network (SNN) for attitude estimation and control of a quadrotor. The network is trained by imitation learning from a reference complementary-filter-plus-PID stack, split into an estimation sub-network and a control sub-network that are merged after training, and deployed on a Crazyflie with a Teensy 4.0 microcontroller at 500 Hz. The authors report stable position tracking and attitude step responses on real flights, with a headline RMSE of 3.03 degrees versus 2.67 degrees for the regular PID stack. They also propose time-shifting training targets, augmenting the training data with disturbances and SNN-in-the-loop flight, and fixing a small subgroup of integrator neurons to stabilize training. The paper concludes with an operations-based energy analysis.","tokens_in":13290,"tokens_out":6924,"duration_ms":70442,"significance":"If the central claims are fully validated, this is a significant step: it would be the first demonstration of a spiking network mapping raw IMU data directly to motor commands for attitude control of a real quadrotor, with stable flight, and with code and training data publicly released. Strong points include the real-robot experiments (ten runs per condition), the modular estimation/control architecture, the practical integrator-neuron mechanism, and the reproducible code repository. However, the quantitative comparison to the PID stack rests on an unvalidated internal attitude estimate and is reported without error bars or significance tests, so the strength of the 'estimation and control' claim is currently not supported by the measurements as presented.","major_comments":[{"comment":"The headline RMSE comparison (3.03 degrees for the SNN vs. 2.67 degrees for the PID) is computed between the commanded roll setpoint and the 'resulting (estimated) roll angle'. For the SNN, this estimated roll is an internal, unsupervised intermediate representation: Section II-C.3 explains that the estimation sub-network has no access to control commands and is trained only through the merged control loss, and Section II-D states that OptiTrack provides only position and absolute heading, not full attitude. The paper therefore never checks the SNN's attitude estimate against ground truth. The internal variable could be a control-relevant latent that is not a physical attitude, in which case the reported RMSE does not measure attitude tracking error. Since the central claim is 'attitude estimation and control', the authors must either validate the estimate against true attitude (e.g., full-pose motion capture or a separate IMU-based ground-truth estimator) or explicitly narrow the claim to end-to-end control without asserting that the internal signal is an attitude estimate.","section":"Section III-B, Table I"},{"comment":"The comparison between controllers reports a single RMSE value per condition with no error bars, confidence intervals, or significance test, despite ten runs per controller. The 'average SD' in Table I is the standard deviation of the response traces across runs, not the variability of the RMSE; it says little about whether the 3.03 vs. 2.67 degree difference, or the differences among SNN variants (3.03, 3.10, 3.24, 3.14 degrees), are meaningful. The authors should report the per-run RMSE distribution (e.g., mean and standard deviation over the ten runs, or box plots) and, if appropriate, a paired test between the SNN and the PID stack. Without this, the abstract's quantitative comparison is not statistically supported.","section":"Table I and Section III-B"},{"comment":"The manuscript repeatedly calls the system 'fully neuromorphic' (abstract and Section I). The experimental implementation runs on a Teensy 4.0, a conventional ARM microcontroller, as stated in Section II-D, and Section III-C explicitly acknowledges that 'The SNN in this research runs on a conventional microprocessor.' The contribution is a spiking algorithm deployed on non-neuromorphic hardware, which is valuable, but calling it a 'fully neuromorphic system' overstates the physical implementation. The authors should rephrase to indicate an all-SNN control pipeline executed on a conventional microcontroller, and reserve 'neuromorphic' for the algorithm or for hypothetical future hardware deployment.","section":"Abstract, Section I, Section II-D, Section III-C"}],"minor_comments":[{"comment":"The cross-correlation analysis that selects the time shift of approximately 6 steps should state explicitly whether it was performed on the training set only or on a held-out validation set, since this affects the risk of overfitting the reported flight results.","section":"Section II-C.1"},{"comment":"The abstract and introduction say the network maps 'raw sensory input directly to motor commands', but the network also receives attitude setpoints as inputs (Section II-B.2). This should be clarified so that 'raw' means 'without an explicit attitude estimator' rather than 'only sensor data'.","section":"Section II-C.2"},{"comment":"The rise-time (RT) metric is not defined in the text. Please specify the threshold used (e.g., 10% to 90% of the steady-state roll change) so the values in Table I are interpretable.","section":"Table I"},{"comment":"In the sentence 'a small delay in the derivative command will induce oscillations', the term 'derivative command' should likely be 'derivative component' or 'derivative action' for clarity.","section":"Section II-C.1"},{"comment":"The energy comparison relies on several assumptions (e.g., the 37x ratio between multiplication and addition energy, and the representation of the complementary filter cost as 15-30 multiplications). These estimates should be stated more cautiously, since the actual energy consumption depends on hardware-specific factors and the authors do not measure power directly.","section":"Section III-C"},{"comment":"There is a typo in the conclusion: 'implicitely' should be 'implicitly'.","section":"Section IV"}],"recommendation":"major_revision","confidential_remarks":"The core flight demonstration is credible and the code/data release is a real asset. The main risk is that the 'attitude estimation' claim is not backed by any ground-truth check of the internal estimate; this needs to be addressed with new experiments or a careful re-scoping. The statistical treatment of the headline RMSE should also be strengthened. The overclaim about 'fully neuromorphic' is fixable by rewording. I would not reject: the contribution is potentially significant and the issues are addressable within a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know before anything else. First, this is the first demonstration I know of where a trained spiking network closes the attitude loop on a real quadrotor, taking raw IMU data to motor commands at 500 Hz and flying stably. Second, the headline quantitative claim—3.03 degrees versus 2.67 degrees for the PID stack—is weaker than it looks, because both numbers are computed against each controller's own internal attitude estimate, and the SNN's estimate is never checked against ground truth. The drone flies, but the \"estimation and control\" headline is partially unproven.\n\nWhat is genuinely new: the modular design of training an estimation sub-network and a control sub-network separately, then merging them by multiplying weight matrices, is a clean idea others will borrow. The practical fixes are also real: hand-fixed integrator neurons to get integral action, time-shifting the training targets to compensate for SNN latency, and data augmentation with injected disturbances. These are described with enough detail to reproduce, and the code and data are public. The real-robot position control results show stable, repeatable flight, which is the core feasibility claim and it holds up.\n\nThe soft spots are real but not fatal. The RMSE comparison lacks error bars and significance tests. The time-shift hyperparameter is chosen from the cross-correlation peak in the same data used for training, which is mildly circular. And calling the system \"fully neuromorphic\" is an overstatement; the network runs on a Teensy microcontroller, not a neuromorphic chip. The authors acknowledge this in Section III-C, so it is rhetorical excess rather than deception.\n\nThe biggest issue is the unattested attitude estimate. The estimation sub-network is trained only through the merged control loss, and the only external references are position and heading. The internal \"roll angle\" could be a control-relevant latent variable that happens to work for the training maneuvers. To support the estimation claim, they should compare the internal estimate to OptiTrack ground truth on attitude, or at least narrow the claim to end-to-end control until that check is done. This is a missing experiment, not a fundamental flaw.\n\nThis paper is for anyone in neuromorphic robotics or imitation learning for low-level control. It deserves a serious referee. I would accept it for peer review and ask for a ground-truth attitude check and proper statistics before final acceptance.","headline":"A real, reproducible demonstration of end-to-end spiking attitude control on a quadrotor, but the estimation claim is under-validated and the headline RMSE comparison scores each controller on its own internal estimate.","tokens_in":13886,"tokens_out":2457,"would_cite":true,"duration_ms":24335,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A spiking neural network trained by imitation can take raw inertial measurements from a real quadrotor and output motor commands directly, tracking attitude almost as well as the standard PID stack.","keywords":["spiking neural networks","neuromorphic control","attitude estimation","imitation learning","quadrotor","end-to-end control","leaky-integrate-and-fire","Crazyflie"],"falsifier":"Run the deployed network through sustained large-angle attitude changes and rapid yaw reversals that are underrepresented in the training set while recording true orientation with an external motion-capture system; if the network's internal attitude readout deviates substantially from the motion-capture orientation even when the motor commands still appear reasonable, then the system is not performing generalizable attitude estimation.","tokens_in":12880,"feed_emoji":"🚁","tokens_out":11782,"duration_ms":102150,"temperature":0.7,"pith_summary":"This paper sets out to show that a spiking neural network can replace the entire low-level attitude loop of a real quadrotor: raw gyroscope and accelerometer readings go in, motor torque commands come out, with no classical estimator or PID controller in between. The authors train two spiking sub-networks separately by imitation learning—one that estimates attitude from sensor data and one that turns attitude into motor commands—and then merge them into a single network. Deployed on a 35-gram Crazyflie with an added microcontroller, the merged network runs at 500 hertz and tracks attitude commands with an average error of 3.03 degrees, compared with 2.67 degrees for the conventional flight stack. What matters for future autonomous drones is that the full perception-to-control pipeline is expressed in spiking neurons, a necessary step toward running all of a drone's autonomy on one low-power neuromorphic chip.","feed_headline":"Flies a real quadrotor with a pure spiking network","feed_subtitle":"Attitude error 3.0° vs 2.7° for the standard PID stack, at 500 Hz on a 35 g drone.","key_machinery":"The central object is a merged spiking neural network built from current-based leaky-integrate-and-fire (CUBA-LIF) neurons—neurons that keep a synaptic current and a membrane potential and emit a binary spike when the potential crosses a threshold. The architecture has two parts: a two-layer recurrent estimation sub-network that converts gyroscope and accelerometer readings into an internal attitude estimate, and a one-layer recurrent control sub-network that maps that estimate plus attitude setpoints into torque commands. Because both sub-networks connect to the outside world through linear weight matrices, the output weights of the estimator and the input weights of the controller are multiplied during merging, yielding one network with no classical estimator or PID in the loop. The mechanism that gives the network integral action is a small group of ten control-layer neurons whose leak and threshold parameters are frozen at 1, turning them into lossless integrators; the mechanism that removes the SNN's phase lag is training the control sub-network on targets shifted roughly six timesteps into the future. Together these mechanisms carry the argument that a spiking network can absorb the temporal operations—integration, differentiation, prediction—that a conventional flight stack performs with explicit filters and controller terms.","core_discovery":"On its own terms, the paper's central claim is that a fully spiking network can perform attitude estimation and control end-to-end on a real quadrotor. The network's inputs are the six raw IMU signals; its outputs are torque commands fed straight into the motor mixer. The authors train an estimation sub-network and a control sub-network separately, then merge them by multiplying their linear input and output weight matrices, so that deployment uses a single network with roughly 15 percent spiking activity. To make the approach work in flight, they add three ingredients to the imitation-learning procedure: ten neurons in the control layer have their leak and threshold parameters fixed to act as perfect integrators, absorbing gyroscope bias the way a PID integral term would; the training targets are shifted about six timesteps into the future so the network learns to predict the expert's next control action, compensating for the SNN's intrinsic delay; and the dataset is augmented with excitation and random disturbances so the network sees states beyond the stable hover regime. With these modifications, the SNN tracks commanded roll angles with 3.03 degrees RMSE (average standard deviation 0.77 degrees), while the PID/complementary-filter stack achieves 2.67 degrees RMSE. The authors read this as evidence that neuromorphic end-to-end control is feasible on current hardware.","pith_inferences":["If the estimation sub-network indeed generalizes, the same modular estimation/control merge could be applied to other low-level loops—for example, optical-flow velocity estimation or wheel-encoder odometry—where a classical observer and a controller are both present.","The power comparison in the paper is based on operation counts rather than measured hardware, so we infer that the real payoff, and the most direct test of the approach, will come when the identical network runs on an actual neuromorphic chip instead of a general-purpose microcontroller.","The authors never compare the internal attitude estimate to external ground truth, so we regard a hidden-layer readout that matches a motion-capture orientation across aggressive maneuvers as the key confirmation that the estimation claim is real.","The time-shift trick suggests a general rule for spiking controllers: because leaky-integrate-and-fire dynamics intrinsically smooth and delay signals, training a network to predict the future reference output is a way to recover responsiveness without adding explicit latency compensation."],"forward_implications":["A single SNN can take over the low-level attitude loop on a real quadrotor, with tracking error within about 0.4 degrees of a conventional complementary-filter-plus-PID stack.","The train-then-merge recipe lets estimation and control be learned separately, which avoids the local minimum in which the estimator ignores sensor data and simply copies the command signal.","Training on time-shifted targets removes most of the SNN-induced delay and reduces oscillations, at the cost of roughly 15 milliseconds of additional rise time.","Fixed-parameter integrator neurons provide bias-rejecting integral action without the unstable recurrent gains that free training tends to produce.","Because inference is dominated by additions and spiking activity is sparse, the network would be substantially cheaper to run on native neuromorphic hardware than on the conventional microcontroller used in the experiments."],"supporting_citations":[{"why":"It supplies the linear weight-matrix merging procedure that lets the separately trained estimation and control sub-networks be collapsed into one network.","marker":"[29]"},{"why":"It defines the Crazyflie platform and its default complementary-filter and cascaded-PID expert policy that provides the imitation-learning targets.","marker":"[32]"},{"why":"It provides the surrogate-gradient method that makes backpropagation-through-time training of the spiking networks possible.","marker":"[34]"},{"why":"It frames the state-distribution mismatch in imitation learning that motivates the dataset augmentation with excitation and disturbances.","marker":"[37]"},{"why":"It motivates structuring recurrent network parameters before training, which leads to the fixed-parameter integrator neurons used for integral action.","marker":"[39]"},{"why":"It supplies the energy-per-operation comparison between integer addition and floating-point multiplication used in the power-consumption analysis.","marker":"[44]"}],"fun_headline_variants":["First neuromorphic autopilot on a real quadrotor","Pure spiking network controls real quadrotor attitude","SNN flies real quadrotor with 3.0° attitude error","Neuromorphic end-to-end control: SNN on Crazyflie","All-spiking attitude control on a real drone"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim assumes that the estimation sub-network, which is merged with the control sub-network and never validated against external ground truth on the real drone, actually computes a generalizable attitude estimate rather than a mapping that only works for the maneuvers and disturbances seen during training.","fun_headline_variants_meta":{"raw":{"variants":["First neuromorphic autopilot on a real quadrotor","Pure spiking network controls real quadrotor attitude","SNN flies real quadrotor with 3.0° attitude error","Neuromorphic end-to-end control: SNN on Crazyflie","All-spiking attitude control on a real drone"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000767,"raw_usage":{"total_tokens":3484,"prompt_tokens":1112,"completion_tokens":2372,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":728,"completion_tokens_details":{"reasoning_tokens":2286}},"tokens_in":728,"tokens_out":2372,"duration_ms":17315,"temperature":1.0,"reasoning_tokens":2286,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:42:44.455966+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the deployed network through sustained large-angle attitude changes and rapid yaw reversals that are underrepresented in the training set while recording true orientation with an external motion-capture system; if the network's internal attitude readout deviates substantially from the motion-capture orientation even when the motor commands still appear reasonable, then the system is not performing generalizable attitude estimation.","supporting_citations":[{"cited_title":"Crazyflie 2.0 quadrotor as a platform for research and education in robotics and control engineering,","cited_arxiv_id":null,"evidence_quote":"It defines the Crazyflie platform and its default complementary-filter and cascaded-PID expert policy that provides the imitation-learning targets."},{"cited_title":"A reduction of imitation learning and structured prediction to no-regret online learning,","cited_arxiv_id":null,"evidence_quote":"It frames the state-distribution mismatch in imitation learning that motivates the dataset augmentation with excitation and disturbances."}],"review_version":1}