{"id":"5f38eae8-488e-4535-b198-a53e37664eda","arxiv_id":"2509.14126","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A decentralized reinforcement learning controller with direct motor commands lets teams of drones carry cable-suspended payloads, recover from harsh disturbances, and transfer from simulation to real Crazyflie hardware.","lead":"The authors trained drone teams to carry a hanging payload using reinforcement learning, with each drone directly commanding its motors and handling cases where the cables go slack or tight. The work shows such learned teamwork can beat a classical planning controller in simulation and transfer to real drones under wind.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Zero-shot sim-to-real claim depends on an untested linear thrust-to-PWM mapping without battery voltage compensation.","rationale":"I agree with the reader that the thrust-to-PWM equivalence is the weakest link in the central sim-to-real claim. A direct measurement would settle it. The baseline mismatch is also serious - the abstract's 'classical decentralized controllers' are not actually tested - but that is an overstatement that can be corrected by rewording and adding baselines. The thrust assumption, by contrast, is a physical premise that could invalidate the transfer claim if wrong. The paper's qualitative hardware evidence is not sufficient to rule out a mapping error, because a closed-loop policy might still fly while underperforming. Therefore the conditional verdict is appropriate, and the proposed thrust-curve measurement is the concrete check that would move the assessment.","tokens_in":12111,"tokens_out":11150,"duration_ms":99109,"concrete_test":"Mount a Crazyflie 2.1 on a thrust stand and measure static thrust versus PWM duty cycle over the full range (e.g., 10% to 100% in 5% steps) at three battery voltages (4.2V, 3.8V, 3.5V). Compare the measured curves to the linear model f = u * f_max used in Section III.C. If any point deviates by more than 10% of max thrust, or if the normalized curve shifts with voltage, retrain the policy with a fitted nonlinear motor model and repeat the zero-shot hardware flights. If the retrained policy's hardware performance drops below the reported qualitative success, the linear mapping is load-bearing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The zero-shot transfer claim rests on the Section III.C assertion that 'in the operating range of the micro brushed motors, thrust is approximately proportional to duty cycle' and on interpreting the same normalized action u_i as a thrust command in simulation and as a PWM duty cycle on the robot. This mapping is not validated by any static thrust measurement, and it is not covered by the domain randomization, which only perturbs per-motor thrust caps and lag time constants. Brushed micro motors exhibit a roughly quadratic thrust-to-duty relationship and a deadband, and battery discharge shifts the curve; the paper even states that PWM is sent 'without battery voltage compensation.' Since the policy operates near actuation limits, the linear assumption may be violated exactly in the regimes the paper emphasizes. The hardware demonstrations are qualitative, so they do not quantify whether the mapping error degrades performance. If the true mapping is nonlinear, the trained policy's motor commands are systematically biased, and the claimed zero-shot transfer is not robust.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CrazyMARL, a decentralized multi-agent reinforcement learning framework for cooperative aerial transport of cable-suspended payloads by multiple quadrotors. The policy is trained with IPPO and shared parameters in a GPU-parallelized MJX simulator, uses a hybrid slack-taut cable model, and outputs direct PWM motor commands at 250 Hz without a low-level cascade. The authors report simulation results with 79.7% recovery from harsh initial conditions versus 43.5% for a baseline, qualitative robustness tests under wind and pushes, and zero-shot sim-to-real transfer on Crazyflie 2.1 hardware. The paper also presents generalization sweeps and scalability results for Q=1,2,3,6 quadrotors, and releases code and videos.","tokens_in":12231,"tokens_out":2884,"duration_ms":29521,"significance":"If the claims hold, this is a meaningful step for decentralized motor-level control of multi-UAV cable-suspended payload transport, especially because it explicitly models slack-taut cable transitions, which many prior rigid-link approaches avoid. The high-throughput training pipeline (2e9 environment steps in 70 minutes on a single GPU), the fully decentralized execution, and the public release of code and videos are concrete strengths that support reproducibility. The real-hardware demonstration is valuable even if qualitative. However, the central performance claims currently rest on a single, mismatched baseline and an unvalidated thrust-to-PWM mapping in the sim-to-real chain, so the significance is conditional on fixing these points.","major_comments":[{"comment":"The abstract claims that the learned policies outperform 'classical decentralized controllers,' but the only baseline used in Section IV.A is the centralized trajectory-optimization method of [7], which the text itself describes as relying on centralized planning with an online tracker and a rigid-rod cable model. A centralized planner is not a decentralized controller, so the experiments do not support the abstract's comparative claim against decentralized methods. Please either add a decentralized model-based baseline (e.g., a decentralized NMPC from [12]) or revise the abstract and conclusion to state that the comparison is against a centralized rigid-rod baseline.","section":"Abstract and Section IV.A"},{"comment":"The zero-shot sim-to-real claim depends on the assertion in Section III.C that 'in the operating range of the micro brushed motors, thrust is approximately proportional to duty cycle,' combined with interpreting the same normalized action as a thrust command in simulation and as a PWM duty cycle on hardware. No static thrust measurements, deadband characterization, or battery-voltage compensation are reported, and the domain randomization covers per-motor thrust caps and lag time constants but not a nonlinear thrust-to-PWM relationship. Since the paper emphasizes operation near actuation limits, this mapping error could be largest exactly in the regimes of interest. The hardware demonstrations are qualitative and therefore do not quantify any degradation from this mapping. Please add static thrust calibration data or battery-voltage compensation, and report quantitative real-world tracking or recovery metrics.","section":"Section III.C and Section IV.D"},{"comment":"The central quantitative comparison reports success rates (797/1000 versus 435/1000) but the text does not define the success criterion beyond 'recover within 10 s.' The reward function in Appendix A includes terms for tracking error, stability, taut cables, and safe spacing, which overlap heavily with the likely definition of recovery, so the evaluation metric must be stated independently and precisely (e.g., threshold on payload position error, settling time, and cable-tautness conditions). Without a clear, reward-independent success definition, the 79.7% versus 43.5% result is difficult to interpret.","section":"Section IV.A and Appendix A"},{"comment":"The generalization claim in Section IV.B is based on sweeps over cable length, payload mass, observation noise, and seed, but the text does not report confidence intervals or trial-level variance for the success rates in Figure 4, even though each point is described as 1000 runs. For a stochastic evaluation, reporting only point estimates makes it impossible to judge whether the observed differences (e.g., the slight drop for lighter payloads) are meaningful. Please add error bars or confidence intervals.","section":"Section IV.B"}],"minor_comments":[{"comment":"Equation (8) and the appendix define the reward components, but the roles of the scalar constants λyaw, λup, λs, and vmax are described only as 'tuning' parameters; a short sensitivity discussion would help the reader understand how robust the reported results are to these choices.","section":"Section III.F"},{"comment":"In the scalability experiment, the statement that Q=6 'exposes a coordination limit' is useful, but the text does not report the failure modes quantitatively (e.g., fraction of trials ending in collision versus timeout). A breakdown would make the scalability claim more informative.","section":"Section IV.C"},{"comment":"The hardware section states that observations use motion capture for payload and teammate positions. Since the paper emphasizes decentralized onboard execution, it would be helpful to clarify whether the motion-capture data are streamed to each robot (and thus a form of external infrastructure) or whether the policy would run with onboard perception only.","section":"Section IV.D"},{"comment":"There are a few formatting artifacts in the text (e.g., 'V iktor' in the author list and the unnumbered equation in Section III.C), and the text sometimes switches between 'UAV' and 'UA V' spacing. These should be cleaned up.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper is timely and the engineering contribution is real, but the headline comparison and the sim-to-real claim need stronger support before publication. The baseline mismatch and the unvalidated thrust-to-PWM mapping are load-bearing; I do not think they are unfixable, but they require additional experiments (a decentralized baseline or a clearly qualified claim, plus static thrust calibration or voltage compensation). I would also encourage the editor to consider whether the framing as the 'first' multi-UAV RL work with cable mode transitions is appropriately qualified given the very recent [32]."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here is my honest read of CrazyMARL. The real new thing is the specific combination: decentralized MARL with direct PWM output at 250 Hz, trained on hybrid slack-taut cable dynamics, and deployed on Crazyflie hardware. The cited multi-UAV RL papers use rigid or always-taut cables, and most rely on low-level cascades, so this is a genuine increment. The engineering is also solid: a MJX/JAX pipeline that trains 2B steps in 70 minutes on one GPU, shared-parameter IPPO, a clearly documented reward, and released code and videos.\n\nThe simulation core is credible. 797/1000 vs 435/1000 recoveries over randomized trials is a concrete, repeatable result, and the generalization sweeps over cable length, payload mass, observation noise, and seeds give a useful picture of where the policy breaks.\n\nSoft spots, ranked by how much they matter. First, the abstract says the policy outperforms 'classical decentralized controllers,' but the only baseline is a centralized rigid-rod trajectory optimizer from the authors' own group. That mismatch overstates the comparison. Second, the zero-shot sim-to-real claim depends on the Section III.C assertion that thrust is approximately proportional to duty cycle on the brushed micro motors. There is no static thrust measurement, no battery-voltage compensation, and the domain randomization does not vary the shape of that curve. The stress-test note is right: this is a load-bearing premise, and the hardware demos are qualitative, so we don't know how much mapping error the policy actually absorbs. The demos look promising, but 'zero-shot transfer' is not yet quantitatively supported. Third, scalability collapses at Q=6; the authors acknowledge it and attribute it to permutation sensitivity, which is fair, but it limits the scope of the claim.\n\nWho should read this: people working on multi-UAV manipulation, RL for aerial robots, or sim-to-real transfer. It deserves a serious referee. The path forward is not a rewrite; it is cleaning up the baseline description, adding quantitative hardware metrics, and either validating the thrust-to-PWM mapping or softening the transfer claim. I would send it to review, expecting revision.","headline":"Genuinely new MARL + direct-PWM + hybrid-cable pipeline with a credible simulation win; the zero-shot transfer claim is real but currently over-sold.","tokens_in":12805,"tokens_out":3229,"would_cite":true,"duration_ms":28194,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A decentralized multi-agent RL policy, trained on hybrid slack-taut cable dynamics, outperforms a classical rigid-rod controller for multi-UAV cable-suspended payload transport, recovering 80% of harsh trials versus 44% for the baseline…","keywords":["multi-agent reinforcement learning","cable-suspended payload","decentralized control","sim-to-real transfer","domain randomization","direct motor control","quadrotor swarm","hybrid cable dynamics"],"falsifier":"Measure static thrust versus PWM duty cycle on the Crazyflie motors across the battery voltage range (3.7 to 4.2 V) and at different motor temperatures; if the curve deviates by more than a few percent from a straight line through the origin, the zero-shot transfer claim is undermined. A direct behavioral test: run the reported push-recovery trials with a worn or partially discharged battery; if the real-world recovery rate drops sharply while the simulator (which models thrust as proportional to $u_i^t$) still predicts high recovery, the central transfer claim is falsified.","tokens_in":11885,"feed_emoji":"🚁","tokens_out":10354,"duration_ms":82851,"temperature":0.7,"pith_summary":"This paper aims to establish that a team of small quadrotors can learn a fully decentralized control policy that directly outputs motor PWM commands to transport a cable-suspended payload, handling the hybrid dynamics of cables that go slack and taut. The authors report that this learned policy outperforms a classical decentralized controller in simulation, recovering from 80% of harsh randomized trials compared with 44% for the baseline, and tracks a figure-eight trajectory with less error. They also report successful zero-shot transfer of the same policy to real Crazyflie quadrotors, which keep a payload stable under wind and pushes. If correct, this demonstrates that end-to-end multi-agent reinforcement learning can replace cascaded low-level controllers for cooperative aerial manipulation on resource-constrained hardware while respecting cable physics that prior multi-UAV work simplified away.","feed_headline":"Decentralized RL recovers 80% of harsh payload drops","feed_subtitle":"One shared policy, trained in simulation, transfers zero-shot to real drones carrying a cable-suspended load.","key_machinery":"The load-bearing identity is the normalized action mapping: each agent emits $a_i^t \\in [-1,1]^4$, converted to $u_i^t = (a_i^t + 1)/2 \\in [0,1]^4$, which the simulator reads as a thrust fraction of each motor's maximum and the real robot reads as a PWM duty cycle; the paper justifies this by asserting that, in the operating range of the micro brushed motors, thrust is approximately proportional to duty cycle. This same-number-different-physics bridge is what makes zero-shot transfer conceivable. The second piece is the hybrid cable model: each quadrotor is connected to the payload by a tendon of length $L$ that exerts tension only when the distance equals $L$ (taut) and zero otherwise (slack), so the policy learns to handle cable mode transitions rather than assuming a rigid rod. The third is the modular reward $r = r_{\\text{track}}\\, r_{\\text{stable}} + r_{\\text{safe}}$, with bounded exponential shaping terms that couple tracking with stability, add safety incentives for smooth and balanced motor commands, and discourage collisions and saturation. Training uses IPPO with shared parameters and a Gaussian action distribution, over 16,384 parallel environments on a single GPU, with domain randomization applied to initial states, motor thrust caps, actuator time constants, observations, and external wrenches.","core_discovery":"The central claim, stated on the paper's own terms, is that a single decentralized policy trained with Independent Proximal Policy Optimization (IPPO) and shared parameters across agents can map each quadrotor's local observation—payload error and velocity, its own pose, rates, and previous action, plus the relative positions of teammates—directly to normalized motor commands at 250 Hz, with no low-level attitude or rate controller in the loop. In simulation, the policy achieves an 80% recovery rate (797 of 1000 trials) for a two-quadrotor cable-suspended payload task starting from harsh randomized states, versus 43.5% (435 of 1000) for a rigid-rod trajectory-optimization baseline tracking a precomputed polynomial; the learned policy also reaches the goal at more than twice the mean speed (0.58 m/s vs 0.27 m/s). The same policy transfers zero-shot to Crazyflie 2.1 hardware, where two quadrotors carry a 10 g payload through autonomous takeoff, figure-eight tracking, and disturbance recovery under wind measured at 3.5 m/s. The paper attributes this to training on hybrid slack-taut cable dynamics, heavy domain randomization of motor thrust caps, actuator lag, observation noise, and external disturbances, and a reward that couples tracking with stability and safety.","pith_inferences":["If the thrust-to-duty-cycle proportionality degrades under battery sag or motor heating, the zero-shot transfer may not survive long flights; the paper's domain randomization over motor thrust caps does not explicitly randomize the proportionality slope itself, leaving a voltage-dependent nonlinearity as an untested failure mode.","The headline comparison against a rigid-rod baseline is favorable to the learned method because the baseline cannot even represent slack-taut transitions; a slack-aware model-based controller might narrow the gap, though none exists for multi-UAV teams, which is precisely the gap this paper fills.","The modular reward and observation structure could be reparameterized for other payload types (rigid links, nets, or multiple suspended loads), suggesting the framework is a general template for decentralized aerial manipulation rather than a single-task solution.","A testable extension: train the same observation and reward on teams of four to eight quadrotors with attention-based or sorted peer encodings; if recovery success degrades gracefully with team size, the permutation-sensitivity diagnosis is confirmed."],"forward_implications":["Multi-UAV teams can transport suspended payloads through slack-taut cable transitions, enabling maneuvers (e.g., recovery from dropped payloads, wide payload swings) that rigid-rod controllers cannot represent.","Direct PWM control at 250 Hz removes the need for cascaded attitude/rate controllers and thrust mixing, allowing the same policy to operate near motor saturation on platforms with low thrust-to-weight ratio.","Because the policy is decentralized and communication-free, teams scale without a central coordinator or inter-robot bandwidth, and the failure of one drone does not collapse the whole control loop.","Domain randomization on motor thrust caps, actuator lag, observation noise, and external disturbances appears sufficient for zero-shot sim-to-real transfer of low-level multi-robot policies, reducing the need for hardware-in-the-loop tuning.","Scaling to larger teams is currently the main limitation: recovery success falls from 81% for two quadrotors to 60% for three and near zero for six, and the paper attributes this to permutation sensitivity in how peer observations are ordered, pointing to order-invariant encodings as the next step."],"supporting_citations":[{"why":"Supplies the classical decentralized baseline (rigid-rod trajectory optimization with an online tracker) that the learned policy is benchmarked against; the 80% vs 43.5% recovery comparison is defined against this method.","marker":"[7]"},{"why":"Defines IPPO, the decentralized training algorithm with shared parameters that produces the policies.","marker":"[25]"},{"why":"Provides the physics simulation engine that models the taut-slack tendon cables and contact-rich dynamics.","marker":"[33]"},{"why":"Supplies the GPU-parallel multi-agent reinforcement learning implementation that enables 16,384 parallel environments and 2 billion training steps.","marker":"[34]"},{"why":"Introduces the first-order actuation lag model on a rotor-speed proxy and the methodology for transferring low-level quadrotor control policies to multiple real robots.","marker":"[35]"},{"why":"The closest prior decentralized MARL approach for cable-suspended loads using a low-level controller and rigid-rod assumption; the paper's improvements (IPPO, direct PWM, hybrid cables) are contrasted against it.","marker":"[32]"}],"fun_headline_variants":["One shared policy flies real drones, recovers 80% of hard payload drops","Decentralized RL beats baseline, recovers 80% in harsh cable-payload tasks","CrazyMARL: Zero-shot sim-to-real for cable-suspended multi-UAV transport","RL drones handle slack-taut cables, zero-shot to real hardware"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The zero-shot transfer rests on the assumption that thrust is approximately proportional to PWM duty cycle for the Crazyflie's motors, so the policy's normalized command $u_i^t$ means the same physical thing in simulation and on hardware; if battery voltage, motor wear, or nonlinearity breaks that proportionality, the transferred policy may fail.","fun_headline_variants_meta":{"raw":{"variants":["One shared policy flies real drones, recovers 80% of hard payload drops","Decentralized RL beats baseline, recovers 80% in harsh cable-payload tasks","CrazyMARL: Zero-shot sim-to-real for cable-suspended multi-UAV transport","RL drones handle slack-taut cables, zero-shot to real hardware"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000503,"raw_usage":{"total_tokens":2508,"prompt_tokens":1046,"completion_tokens":1462,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":662,"completion_tokens_details":{"reasoning_tokens":1374}},"tokens_in":662,"tokens_out":1462,"duration_ms":10701,"temperature":1.0,"reasoning_tokens":1374,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:51:50.817082+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure static thrust versus PWM duty cycle on the Crazyflie motors across the battery voltage range (3.7 to 4.2 V) and at different motor temperatures; if the curve deviates by more than a few percent from a straight line through the origin, the zero-shot transfer claim is undermined. A direct behavioral test: run the reported push-recovery trials with a worn or partially discharged battery; if the real-world recovery rate drops sharply while the simulator (which models thrust as proportional to $u_i^t$) still predicts high recovery, the central transfer claim is falsified.","supporting_citations":[{"cited_title":"Kinodynamic motion planning for a team of multirotors transporting a cable- suspended payload in cluttered environments,","cited_arxiv_id":null,"evidence_quote":"Supplies the classical decentralized baseline (rigid-rod trajectory optimization with an online tracker) that the learned policy is benchmarked against; the 80% vs 43.5% recovery comparison is defined against this method."},{"cited_title":"Jaxmarl: Multi-agent rl environments and algorithms in jax,","cited_arxiv_id":null,"evidence_quote":"Supplies the GPU-parallel multi-agent reinforcement learning implementation that enables 16,384 parallel environments and 2 billion training steps."},{"cited_title":"Sim-to-(multi)-real: Transfer of low-level robust control policies to multiple quadrotors,","cited_arxiv_id":null,"evidence_quote":"Introduces the first-order actuation lag model on a rotor-speed proxy and the methodology for transferring low-level quadrotor control policies to multiple real robots."}],"review_version":1}