{"id":"041c86b4-238a-40c5-9f11-66c1bd187ac3","arxiv_id":"2507.23339","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A reinforcement learning policy trained in a GPU simulator with domain randomization transfers directly to a physical 1/10-scale individual-wheel-drive RC car, sustaining controlled drifts on circular, eight-shaped, and variable-curvature paths.","lead":"This paper trains a reinforcement learning controller that makes a small four-wheel-drive RC car drift along curved paths, using a fast simulator with randomized conditions, and shows it works on the real car. It is a readable recipe for getting simulated driving policies to transfer to physical hardware without retuning.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Real-world variable-curvature results show the policy using a wheel-speed strategy opposite to its simulation behavior and producing ~3x larger tracking error; the central 'effective sim-to-real transfer' claim is therefore not supported by the quantitative evidence.","rationale":"The reader's weakest assumption identifies model fidelity as the key risk; I partially agree but sharpen it. The paper itself provides internal evidence of model mismatch: in Section V.D the real variable-curvature run uses a wheel-speed distribution opposite to simulation, and Table II shows real-world tracking error roughly three times larger. This is not a hypothetical sensitivity to unmodeled dynamics; it is a measured divergence in the control strategy actually deployed. The load-bearing requirement for the central claim is not merely that some policy can drift on the real car, but that the simulation-trained policy transfers without fine-tuning in a way that preserves the learned behavior. The observed qualitative inversion means the policy is operating in a different regime on hardware, so the simulator has not captured the physical drift equilibrium. The concrete test—replaying the real initial condition through the frozen policy in the simulator—would settle whether the mismatch is due to missing lateral load transfer (which the paper itself flags as future work) or to benign multi-modality of the reward landscape. If the simulator fails to reproduce the real trajectory, the no-fine-tuning transfer claim is not established, and the paper should be revised to present the real-world result as robust policy deployment rather than faithful sim-to-real transfer, or add a quantitative state/action comparison and a real-world baseline. This is consistent with the reader's CONDITIONAL verdict, so no verdict change is proposed.","tokens_in":12051,"tokens_out":5150,"duration_ms":63242,"concrete_test":"Use the motion-capture initial state and reference trajectory from the real variable-curvature run to initialize the paper's GPU simulator, then run the frozen policy in closed-loop simulation. If the simulated trajectory retains the right>left wheel-speed asymmetry and approximately -50 degree sideslip while the real run exhibited left>right asymmetry and smaller sideslip, the simulator's dynamics do not replicate the physical car, directly invalidating the 'no fine-tuning' sim-to-real transfer claim. If the simulation does reproduce the real behavior, then the opposite strategy is a policy non-uniqueness and the concern is vacated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section V.D reports that on the variable-curvature track the simulated policy maintains drift with higher right wheel speeds, while the physical Xcar achieves drift with higher left wheel speeds, 'with the compromise of smaller sideslip angles' (Fig. 8b/e, Table II). The position RMSE is 0.231 m real vs 0.075 m simulated (3.1x), and the eight-shaped path shows 0.221 m vs 0.138 m. The central claim—that DR-based RL bridges the sim-to-real gap so the policy transfers without fine-tuning—requires that the trained policy's behavior in simulation is a reliable predictor of real-world behavior. The observed qualitative inversion of the actuation strategy and the degraded tracking show the opposite: the policy still drifts, but it discovers a different, more conservative regime on the real car. This is consistent with the admitted omission of lateral load transfer and suspension dynamics (Section VI), which are first-order effects at the reported sideslip angles. Because the paper provides no real-world baseline and no quantitative state/action comparison between sim and real, the 'effective transfer' statement is not actually demonstrated; it is only shown that a robust policy can drift on the physical car, not that the simulation faithfully captured the dynamics that produce that drift.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a reinforcement learning framework for autonomous drifting on a custom 1/10-scale Individual Wheel Drive (IWD) RC car. The framework combines a simplified planar vehicle model with Pacejka tires, GPU-accelerated parallel simulation, PPO training, and domain randomization over trajectories, initial states, tire parameters, and an autoregressive disturbance process. The authors validate the approach in simulation on circular, eight-shaped, and variable-curvature paths, and on the physical Xcar platform for the eight-shaped and variable-curvature paths, with an additional Olympic-rings demonstration. The central claim is that the framework achieves effective sim-to-real transfer without real-world fine-tuning, enabling precise trajectory tracking with controlled sideslip angles in both simulation and on hardware.","tokens_in":12319,"tokens_out":5704,"duration_ms":66974,"significance":"If the central claim is fully supported, the paper would offer a valuable and reproducible recipe for learning drifting policies in simulation and deploying them directly on physical IWD vehicles. Strengths include the open-sourced hardware and code, real-world experiments with video documentation, a fast GPU-based training pipeline, and a domain-randomization ablation study. The paper is honest about some limitations, such as the omission of suspension and lateral load transfer. However, the quantitative evidence for faithful sim-to-real transfer is incomplete: the real-world variable-curvature behavior differs qualitatively from simulation in the actuation strategy, sideslip magnitude, and tracking error. The contribution is still useful, but the central claim needs either additional evidence or more careful wording.","major_comments":[{"comment":"The central claim of effective sim-to-real transfer is not fully supported by the reported behavioral comparison. On the variable-curvature track, the physical vehicle maintains drift with higher left wheel speeds while the simulated policy uses higher right wheel speeds; the real sideslip angles are smaller, and the position RMSE is 0.231 m versus 0.075 m in simulation, a 3.1x gap. The paper itself describes this as an 'opposite strategy' and a 'compromise,' which is inconsistent with the earlier phrase 'remarkable consistency.' To support the transfer claim, the paper should provide a quantitative comparison of state and action trajectories (for example, distributions of wheel-speed differentials, sideslip, and yaw rate), explain the actuation inversion by identifying the responsible model mismatch, and ideally include a real-world baseline such as a model-based controller or a policy trained without the full domain randomization.","section":"Section V.D, Table II, Fig. 8"},{"comment":"The simplified planar model omits lateral load transfer, suspension, and motor/ESC dynamics; at the reported sideslip angles around 50 degrees these can be first-order effects, and the paper lists them as future work. Domain randomization is the main mechanism claimed to compensate for these omissions, but the only ablation evidence is simulated. Table III reports success rates and RMSE in simulation, and the statement that 'only the full randomization configuration successfully transfers to real-world deployment' is not accompanied by any real-world ablation data. The authors should either provide real-world comparisons with and without key randomization components, or explicitly restrict the transfer claim to the demonstrated cases.","section":"Section III-B, III-C, IV-C, Appendix Table III"},{"comment":"The reward includes r_drift = -(beta - beta_ref)^2, but beta_ref is never defined or given a numerical value. Section IV-C.2 says only that the desired sideslip direction is determined by local curvature. Because this reward term directly shapes the learned sideslip behavior, the reported natural drift equilibrium and the real-world sideslip magnitudes are not reproducible without knowing beta_ref and how it varies along the reference path. Please specify the formula or value, and state whether beta_ref is fixed or curvature-dependent.","section":"Section IV-B.3, Eq. (7), Section IV-C.2"}],"minor_comments":[{"comment":"The text says the simulator handles '106 car instances'; the exponent appears to be missing and should read 10^6 car instances.","section":"Section III-C"},{"comment":"The reward term r_speed = min(0, V - 0.5) penalizes only speeds below 0.5 m/s and does not incentivize the 1.5-2.5 m/s operating range described in the text; please clarify the wording or adjust the reward.","section":"Section IV-B.3, Eq. (12)"},{"comment":"The wheel longitudinal velocities v_ij^x used in the slip penalty are not defined; for a rigid body they depend on the yaw rate and wheel position, so please specify the kinematic relationship.","section":"Section IV-B.3, Eq. (11)"},{"comment":"The text says removing initial-state randomization causes a '29% drop' in success rate; from 75.0% to 46.0% this is a 29 percentage-point drop, not a 29% relative drop, so please use precise terminology.","section":"Appendix, Table III"},{"comment":"The phrase 'remarkable consistency' in the first paragraph is difficult to reconcile with the 'opposite strategy' sentence later in the same section; please revise to avoid overstatement.","section":"Section V.D"},{"comment":"The Olympic-rings demonstration is presented without quantitative tracking errors or sideslip time histories; please add summary statistics or label the figure as illustrative.","section":"Section V.D, Fig. 1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of a robotics conference or journal and the open-sourcing of the hardware and code is a genuine asset. The main risk is overclaiming the sim-to-real transfer: the real-world variable-curvature experiment shows qualitatively different actuation and degraded tracking, so the central claim needs to be either better supported with quantitative behavioral comparisons or carefully weakened. I would encourage the editor to request a revision that addresses this gap."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The real news here is that the authors built a 1/10-scale IWD car, trained a PPO drifting policy in a custom GPU simulator in about ten minutes, and deployed it on the physical car without any fine-tuning on two non-trivial paths (eight-shaped and variable-curvature). That is a genuinely useful recipe, and the open-source platform and code make it reproducible. The ablation study is also a plus: it shows initial-state and trajectory randomization matter most, and that full randomization is what transfers. I buy the central claim in the narrow sense that the policy is robust enough to drift on the real car; that is not nothing.\n\nThe soft spots are real but not fatal. First, the paper overstates consistency. On the variable-curvature track the real car uses the opposite wheel-speed strategy from simulation (higher left, not higher right) and the position RMSE is 0.231 m versus 0.075 m simulated — about 3x. The abstract says \"consistent performance,\" which is too strong. The stress-test note is right that this shows the simulation is not faithfully predicting the real behavior; it is wrong, though, to conclude the transfer claim fails. Transfer means the policy was usable, not that sim and real match. Second, there is no real-world baseline, so absolute RMSE values are hard to interpret. Third, the claimed IWD-over-RWD advantage appears only in simulation. Fourth, the model omits lateral load transfer and suspension, which the authors admit; for an RC car this is acceptable but limits the conclusions about generality.\n\nNone of these undermine the paper's core value as an engineering demonstration and a reusable testbed. The main fixes a referee should request are a real-world baseline (even a simple PID or a non-drift racing line), a quantitative sim-vs-real state comparison, and a more careful wording of what \"effective transfer\" means.\n\nWho is this for? Anyone working on sim-to-real RL for agile vehicle control, and people who want a cheap open platform for IWD experiments. It deserves a serious referee. My recommendation: send it out, with revisions likely.","headline":"Solid engineering contribution: an open-source IWD RC platform, fast DR-based RL training, and real drifting without fine-tuning, though the sim-to-real gap is wider than the paper admits.","tokens_in":12845,"tokens_out":2350,"would_cite":true,"duration_ms":28906,"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":"This paper claims that a policy trained only in a domain-randomized simulator can steer a real individual-wheel-drive car through sustained drifting maneuvers on complex paths without any real-world fine-tuning.","keywords":["autonomous drifting","reinforcement learning","domain randomization","sim-to-real transfer","individual wheel drive","trajectory tracking","GPU parallel simulation","vehicle control"],"falsifier":"Run the released policy on the same Xcar platform on a surface with a measurably different friction coefficient, such as wet painted concrete or a polished floor, and record whether it keeps sideslip angle within about ±15 degrees of the reference and position RMSE below 0.5 m on the figure-eight path; the paper's claim of transferability across tire-road conditions would be contradicted if the vehicle spins out or the tracking error exceeds this bound without retraining.","tokens_in":11857,"feed_emoji":"🏎️","tokens_out":7143,"duration_ms":74063,"temperature":0.7,"pith_summary":"The paper claims that a car can learn to drift entirely in simulation and then execute sustained, controlled drifts on real asphalt without any fine-tuning. The training recipe is a reinforcement learning policy (proximal policy optimization) run in a very fast GPU-based simulator with randomized tire parameters, randomized starting states, and an autoregressive disturbance on tire forces. On a 1/10-scale car with four independently driven wheels, the policy tracks circles, figure-eights, and variable-curvature paths while holding sideslip angles around 45 to 55 degrees, with real-world position errors below about 0.23 m. The central assertion is that the combination of high-throughput parallel simulation and systematic domain randomization is what bridges the simulation-to-reality gap, and that independent wheel drive gives extra yaw authority that makes drifting easier to initiate and control than with rear-wheel drive.","feed_headline":"Sim-trained RL drifts real IWD car with no fine-tuning","feed_subtitle":"A domain-randomized, GPU-parallel policy learns 45-55 degree sideslip drifting on circles and figure-eights, then transfers directly.","key_machinery":"The load-bearing mechanism is the Markov decision process design: a rotation- and translation-invariant observation vector ($r$, $\\beta$, $V$ plus wheel velocities and previous controls), path information expressed as future waypoints in the vehicle frame, and error signals ($e_{pos}$, $e_{dir}$, $e_\\kappa$, $e_\\beta$) defined in a curvilinear coordinate system attached to the reference trajectory, letting one policy handle arbitrary paths without overfitting to global coordinates. The other half of the machinery is the domain-randomization stack: uniform randomization of Pacejka tire parameters $B\\in[0.8,1]$, $C\\in[2,2.5]$, $D\\in[0.3,0.4]$, Gaussian perturbations to initial position and heading, and an autoregressive disturbance $d_{t+1}=a d_t + w \\varepsilon_t$ on tire forces, which together make the policy robust to unmodeled dynamics. The GPU-based parallel simulator (about $10^5$ vehicles in parallel, converging in roughly 10.8 minutes on one RTX 3080) is what makes the data-hungry RL training feasible in practice.","core_discovery":"The paper's central claim is that a PPO-trained policy with an observation space built from coordinate-free vehicle states (yaw rate, sideslip, speed) plus future waypoints and curvilinear tracking errors, and an action space of steering plus four independent wheel speeds, learns to drift along prescribed trajectories with large controlled sideslip angles. The key discovery is that zero-shot sim-to-real transfer works when training is done in a massively parallel simulator with domain randomization over tire parameters B, C, D, initial states, and an autoregressive disturbance injected into tire forces. The authors report that while removing dynamic disturbance improves simulated success rate (80 versus 75 percent), only the full randomization configuration transferred to the physical car, so the disturbance term is essential for overcoming unmodeled dynamics. Real-world runs on a figure-eight path and a variable-curvature track reproduce the simulated sideslip and speed profiles, with position RMSE of 0.221 m and 0.231 m respectively, and the vehicle completes multiple drift-direction reversals on an Olympic-rings-inspired path. A comparative rear-wheel-drive experiment shows that the individual-wheel-drive system reaches the drift equilibrium in about 1 second versus 1.5 seconds and with slightly better tracking, evidence for the control-authority benefit of independent wheel drives.","pith_inferences":["The transfer claim is demonstrated on only one physical platform and surface; a natural stress test is whether the same randomization set transfers to a different-scale vehicle or to asphalt with a different friction coefficient, which the authors imply but do not run.","The observed left/right strategy inversion between simulation and reality on the variable-curvature track hints that the simulator does not fully capture asymmetry; a testable extension is to randomize left-right tire properties separately and check whether the policy learns symmetric-in-symmetry, adaptive strategies.","The same domain-randomization plus parallel-simulation recipe may apply to other limit-handling skills beyond drifting, such as emergency obstacle avoidance or driving on low-friction surfaces, since the method does not encode drift-specific structure beyond the reward and path representation."],"forward_implications":["A drifting controller can be obtained in about 11 minutes of GPU time and deployed directly on the vehicle, so iteration on reward and randomization becomes practical for a single researcher with one GPU.","The observed wheel-speed differentials during drift initiation show a concrete torque-vectoring behavior, making the policy a candidate for further study of how IWD authority should be exploited at the friction limit.","The open-source platform and codebase allow other groups to reproduce the exact policy and hardware results, which is the standard for a transferable recipe.","Because the simulator supports RWD and AWD, the same observation, reward, and randomization design can be used to train and compare drift controllers across different drivetrain configurations."],"supporting_citations":[{"why":"Supplies the proximal policy optimization algorithm used to train the drifting policy.","marker":"[33]"},{"why":"Provides the GPU-based parallel simulation paradigm on which the fast simulator is modeled.","marker":"[30]"},{"why":"Prior RL drifting approach requiring over 11 hours of training; forms the training-efficiency baseline.","marker":"[1]"},{"why":"Prior simulation-to-reality failure case for RL drifting on arbitrary trajectories; the gap this paper targets.","marker":"[11]"},{"why":"Early demonstration of RL-based circular drift stabilization on small-scale cars; the feasibility baseline.","marker":"[10]"},{"why":"Basis of the simplified 7-DOF vehicle dynamics model used in the simulator.","marker":"[18]"},{"why":"Additional source for the 7-DOF drifting dynamics and equilibria used in modeling.","marker":"[29]"},{"why":"Supplies the variable-curvature trajectory patterns used in the evaluation.","marker":"[15]"}],"fun_headline_variants":["Zero-shot sim-to-real drifting on IWD RC car","RL drift policy transfers from sim to real IWD car","Domain-randomized RL learns drifting, transfers directly","Sim-trained RL achieves real drifting with no fine-tuning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The assumption that a planar vehicle model with randomized Pacejka tire parameters and an added autoregressive force disturbance adequately captures the real car's unmodeled dynamics (suspension, load transfer, motor and controller response) such that a policy trained only in this simulation remains stable and accurate on the physical vehicle without any fine-tuning.","fun_headline_variants_meta":{"raw":{"variants":["Zero-shot sim-to-real drifting on IWD RC car","RL drift policy transfers from sim to real IWD car","Domain-randomized RL learns drifting, transfers directly","Sim-trained RL achieves real drifting with no fine-tuning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000176,"raw_usage":{"total_tokens":1295,"prompt_tokens":958,"completion_tokens":337,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":574,"completion_tokens_details":{"reasoning_tokens":273}},"tokens_in":574,"tokens_out":337,"duration_ms":4074,"temperature":1.0,"reasoning_tokens":273,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T10:49:46.681171+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the released policy on the same Xcar platform on a surface with a measurably different friction coefficient, such as wet painted concrete or a polished floor, and record whether it keeps sideslip angle within about ±15 degrees of the reference and position RMSE below 0.5 m on the figure-eight path; the paper's claim of transferability across tire-road conditions would be contradicted if the vehicle spins out or the tracking error exceeds this bound without retraining.","supporting_citations":[{"cited_title":"High-speed autonomous drifting with deep reinforcement learning,","cited_arxiv_id":null,"evidence_quote":"Prior RL drifting approach requiring over 11 hours of training; forms the training-efficiency baseline."},{"cited_title":"Deep Drifting: Autonomous Drifting of Arbitrary Trajectories using Deep Reinforcement Learning,","cited_arxiv_id":null,"evidence_quote":"Prior simulation-to-reality failure case for RL drifting on arbitrary trajectories; the gap this paper targets."},{"cited_title":"Autonomous drifting using simulation-aided reinforcement learning,","cited_arxiv_id":null,"evidence_quote":"Early demonstration of RL-based circular drift stabilization on small-scale cars; the feasibility baseline."},{"cited_title":"On steady-state cornering equilibria for wheeled vehicles with drift,","cited_arxiv_id":null,"evidence_quote":"Basis of the simplified 7-DOF vehicle dynamics model used in the simulator."},{"cited_title":"Vehicle drifting dynamics: discovery of new equilibria,","cited_arxiv_id":null,"evidence_quote":"Additional source for the 7-DOF drifting dynamics and equilibria used in modeling."},{"cited_title":"A controller for automated drifting along complex trajectories,","cited_arxiv_id":null,"evidence_quote":"Supplies the variable-curvature trajectory patterns used in the evaluation."}],"review_version":1}