{"id":"a4b98b6d-abde-499b-9d32-94182449bdbd","arxiv_id":"2505.06561","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Reverse curriculum RL lets a simulated quadruped learn to mount a skateboard from nearby starting positions, with qualitative signs of transfer to a mobile board in simulation.","lead":"A simulated Unitree A1 quadruped robot learns to climb onto a skateboard by training backward from the goal state with reverse curriculum reinforcement learning. The work tackles the mounting phase that earlier skateboarding robots skipped, but reports only simulation results with no quantitative success metrics.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Transfer claim is internally inconsistent: 'without additional training' followed by new reward terms and termination condition for the unfixed-board stage; zero-shot transfer is unsubstantiated.","rationale":"After reading the manuscript, the central claim is that reverse curriculum RL produces a policy that mounts a skateboard and that this policy transfers to a mobile board without additional training. The most load-bearing condition is the transfer step, because without it the contribution reduces to mounting on a rigidly fixed board, which is not the stated novelty. The reader's weakest assumption captures exactly this. In my stress-test I found a concrete internal inconsistency that sharpens the concern: Section IV-B says the final training stage allowed the board to move, while Section IV-F simultaneously claims zero-shot transfer and introduces new reward/termination terms 'to further refine the results.' These statements are mutually contradictory under standard RL practice. Modifying the reward changes the training objective; if the policy is not re-trained, the modifications are inert; if it is re-trained, the transfer is not zero-shot. The paper never clarifies which. This is more specific than a generic sim-to-sim generalization concern and it is directly testable from the released code. I concur with the reader's CONDITIONAL verdict: the paper should be accepted only after the authors clarify the final-stage protocol and provide quantitative success statistics for both fixed-board and mobile-board evaluation. Because the reader already reached this verdict, no adjustment is needed. I also note the paper honestly lists real-world validation as future work, so the scope is simulation-only; however, even within simulation, the transfer claim needs the stated clarification.","tokens_in":7840,"tokens_out":6149,"duration_ms":57121,"concrete_test":"Check the training harness in the linked repository. Identify whether the 'final stage' checkpoint is obtained by resuming the fixed-board checkpoint with modified reward weights or by direct evaluation of the fixed-board checkpoint. Concretely, run the repository's evaluation script on the fixed-board checkpoint in the unfixed-board environment with the original reward/termination settings, over at least 100 episodes, recording success rate and mounting time. Then run the training script for the final stage (with overturning penalty and displacement penalty) and compare the resulting policy. If the zero-shot success rate is high (e.g., >90%) and comparable to the refined policy, the transfer claim stands.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV-B states that the skateboard was allowed to move freely 'only in the final stage of training,' implying the policy is trained on a mobile board. Section IV-F, however, claims the fixed-board agent 'demonstrated good performance without requiring additional training,' then immediately introduces a new episode termination condition (skateboard overturning) and negative rewards for overturning and displacement. Changing reward terms and termination conditions alters the optimization objective; for these changes to influence behavior, the policy must be re-trained or fine-tuned, yet the paper does not state that training was resumed. If the final policy used in the reported results was fine-tuned with these modified rewards, the central claim of zero-shot fixed-to-mobile transfer is false. If it was not fine-tuned, the 'refinement' with new rewards is unexplained and the policy was evaluated under a different termination condition than trained. Either way, the current text does not substantiate the transfer claim. The results section (IV-H) provides only qualitative phrasing ('consistent mounting in ~3 seconds') with no success rates, seed variation, or baseline numbers, so this ambiguity is not resolved by numerical evidence.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Reverse Curriculum Reinforcement Learning approach to train a simulated Unitree A1 quadruped to mount a skateboard. The training starts with the robot already positioned on a rigidly fixed skateboard, then gradually expands the spawn distribution around the board, and finally addresses an unfixed (mobile) skateboard. The authors report that the trained policy can mount the skateboard from various starting positions and that a policy trained on a fixed board performs well on a mobile board. The paper includes training curves, a qualitative description of the mounting sequence, and a public repository with code and trained models.","tokens_in":8056,"tokens_out":2444,"duration_ms":26963,"significance":"If the central claims are substantiated, the work would provide a useful demonstration of autonomous skateboard mounting, a skill that prior quadrupedal skateboarding work has not addressed. The main positive elements are the open-source release of code and trained models, the use of a physically plausible skateboard model, and the application of a reverse curriculum to a contact-rich underactuated task. However, the paper's current evidence is largely qualitative, and the transfer claim for the mobile board is internally ambiguous. The contribution is more of an empirical demonstration than a methodological advance, but it could be a valuable data point for the community if the evaluation is tightened.","major_comments":[{"comment":"The transfer claim is internally inconsistent. Section IV-B states that the skateboard was allowed to move freely 'only in the final stage of training,' which implies the final policy was trained on a mobile board. Section IV-F, however, states that the agent trained on a fixed skateboard 'demonstrated good performance without requiring additional training,' and then immediately introduces a new episode termination condition and negative rewards for overturning and displacement. If the final policy was fine-tuned with these modified rewards, the claim of zero-shot transfer is false. If it was not fine-tuned, the added negative rewards could not have influenced the policy, and the evaluation used a different termination condition than the one used during training. The manuscript must clarify exactly which policy was used for the mobile-board results, whether training was resumed, and what role the new termination condition and rewards played.","section":"Section IV-F, Section IV-B"},{"comment":"The results section provides no quantitative success metric. The claim that the robot 'consistently mounts the skateboard in approximately 3 seconds from first contact' is not supported by any tables, success rates, episode counts, or repeated-seed statistics. Figure 2 shows only mean episode reward curves without variance or separate success-rate curves. Since the central contribution is an empirical policy result, the paper needs quantitative evaluation: success rate over a fixed number of evaluation episodes, distribution of mounting times, and variability across training seeds.","section":"Section IV-H, Figure 2"},{"comment":"The forward-curriculum baseline is described only narratively: 'the most promising result observed with this method was successful alignment of the robot's center of mass and coordinate axes with those of the skateboard, along with the placement of two legs on the board.' No reward curves, success rates, or quantitative comparisons are given for the baseline. The claimed advantage of the proposed reverse curriculum over this baseline is therefore not evidenced. The paper should include at least final reward values or success rates for both methods under identical evaluation conditions.","section":"Section IV-E"}],"minor_comments":[{"comment":"Table I lists the supplementary reward terms but does not report their weights or the values of the free parameters sigma, d_th, and k_angular beyond a passing mention in Section IV-B. Since the reward weights directly determine the learned behavior, they should be reported in the text or table even if they are available in the code repository.","section":"Section IV-D, Table I"},{"comment":"The phrase 'The weights of the awards and all other training parameters have not changed' is unclear: it could mean that the reward weights were kept constant across curriculum stages, but the reward terms themselves changed. Please state explicitly which reward terms and weights were active in each curriculum stage.","section":"Section IV-F"},{"comment":"The observation space includes foot contact status with the skateboard, and the authors note that this is difficult to obtain in a real-world setting and is planned to be removed in future work. This is a relevant limitation for the sim-to-real potential of the method and should be acknowledged in the conclusion as well.","section":"Section IV-C"},{"comment":"The training curves in Figure 2 are not labeled with the corresponding curriculum stages on the plot itself, and the axes are not described in the caption. Please make the figure self-contained by marking the stages and defining the reward metric.","section":"Figure 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a plausible empirical demonstration, but the mobile-board transfer claim is currently ambiguous to the point of being unverifiable from the text. If the authors can clarify whether fine-tuning occurred and provide quantitative evaluation with seeds and baselines, the contribution could be acceptable for a workshop or a short conference paper. As it stands, the evidence is too thin for a full archival publication, but the issues are fixable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one genuinely new thing here is the task: prior skateboarding work starts with the robot already on the board, and this paper addresses the missing mounting phase with reverse curriculum RL. That is a real gap and a reasonable method choice. The staged curriculum, from fixed board and robot on top, to spawned nearby, to (claimed) free board, is clearly described. The authors also ship code and trained models and are honest that real-world validation is future work. So this is a useful extension of known techniques to a new contact-rich problem.\n\nThe soft spots are significant, though. The most serious is the transfer claim. Section IV-B says the board was allowed to move freely only in the final stage of training. Section IV-F says the fixed-trained agent 'demonstrated good performance without requiring additional training,' then immediately introduces a new termination condition (board overturning) and negative rewards for overturning and displacement. That is contradictory. If the policy was fine-tuned with those new rewards, the claim of zero-shot transfer is false; if it was not, the new rewards could not have influenced behavior and the policy was evaluated under conditions it never trained on. The stress-test note is right about this. The authors need to state explicitly whether the mobile-board results come from the fixed-board policy as-is or from a fine-tuned policy, and describe the training protocol accordingly.\n\nThe evaluation is also too thin. Figure 2 shows only mean reward curves, with no variance or seed statistics. The success claim in Section IV-H is qualitative: 'consistently mounting in ~3 seconds,' with no success rate, no number of trials, no random initial states table. The forward curriculum baseline is described narratively as having failed, but without numbers or a learning curve. That makes it hard to verify that reverse curriculum is actually what enabled the behavior.\n\nOn the math side, the equations are standard reward shaping and PPO; there is no circular reasoning. The fixed-board mounting itself is plausible—training curves and the sequence of frames suggest something real happened in simulation. The problem is the central transfer claim being underspecified and the lack of quantitative grounding.\n\nWho is this for? Researchers working on quadruped-object interaction and curriculum RL for contact-rich skills. They would get a clear task setup and a starting point, but they should not cite the mobile-board transfer as established fact. With the current text, the paper should go to a serious referee, but it needs major revision: clarify the training protocol for the mobile board, add success rates and seed statistics, and quantify the forward-curriculum comparison. I would send it to review rather than desk-reject, because the task novelty is real and the approach is sensible—but acceptance should be conditional on fixing the transfer claim and the evaluation.","headline":"A genuinely new task (quadruped mounting a skateboard) with a sensible reverse-curriculum approach, but the mobile-board transfer claim is internally inconsistent and the evaluation is too qualitative to support the core claim.","tokens_in":8589,"tokens_out":1382,"would_cite":false,"duration_ms":16203,"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 quadrupedal robot can learn to mount a skateboard by training backward from the goal, and the resulting policy transfers to a moving board.","keywords":["quadrupedal robots","skateboard mounting","reverse curriculum learning","reinforcement learning","proximal policy optimization","contact-rich manipulation","sparse rewards","sim-to-real transfer"],"falsifier":"Run the trained policy on a physical quadruped facing a real skateboard under the same starting offsets used in simulation; if the success rate is near zero because the board shifts during foot placement, the transfer claim fails. A cheaper in-simulation test is to vary skateboard mass, wheel friction, and bushing stiffness beyond the training distribution and watch for success collapse.","tokens_in":7662,"feed_emoji":"🛹","tokens_out":8941,"duration_ms":78691,"temperature":0.7,"pith_summary":"The paper tries to establish that a quadrupedal robot can learn to climb onto a skateboard by itself, using a reinforcement-learning curriculum that begins at the end of the task and works backward. Previous work on robot skateboarding assumed the robot was already standing on the board; the mounting phase, a sparse-reward and contact-rich maneuver, had not been demonstrated. The authors train entirely in simulation with the board fixed, then gradually widen the robot's starting positions, and finally report that the resulting policy performs well on a skateboard free to move without any additional training. If this is right, it fills the missing first stage of an autonomous skateboarding pipeline for legged robots.","feed_headline":"Quadruped learns to mount a skateboard via reverse curriculum","feed_subtitle":"Training starts with the robot on the board, then expands to harder starts, and transfers to a moving board.","key_machinery":"The load-bearing mechanism is reverse curriculum learning: the agent starts each episode in the already-mounted state, learns to stay balanced, then the spawn region is widened in stages, up to a 60 cm by 60 cm area around the board's center of mass, with a final stage that lets the board move freely. The policy is a multilayer perceptron trained with Proximal Policy Optimization, using sparse contact rewards for each foot on the deck, dense distance- and orientation-alignment rewards with exponential kernels, and penalties for board overturning and displacement. The skateboard model treats the truck as a roughly 1:1 map from deck tilt angle $\\alpha$ to wheel steering angle $\\beta$ ($\\beta \\approx \\alpha$) and models the bushings as a torsional spring, which makes the board's response to the robot's weight shift physically reactive.","core_discovery":"The paper's central claim is that a quadrupedal robot can be trained entirely in simulation to mount a skateboard from various starting positions, and that a policy trained with the skateboard rigidly fixed in the world frame performs well when the board is allowed to move freely, with no fine-tuning on the moving board. The authors frame this as closing the gap between walking and riding: earlier skateboarding systems begin with the robot already on the board, leaving the mounting phase open. They report that the trained policy mounts the board in roughly three seconds from first contact, and that a forward curriculum baseline, which spawns the robot near the board and lets it explore freely, never progresses beyond aligning itself and placing two feet on the deck.","pith_inferences":["Beyond the paper: the fixed-to-moving transfer suggests the learned policy relies mostly on board-relative features such as edge points, contact state, and orientation, so the same curriculum may work for other mobile objects whose target configuration is known; the authors do not test this.","Beyond the paper: a real-world transfer test would be the first decisive check, because the simulation's no-slip wheel contact and simplified truck model may be more forgiving than a physical skateboard; the authors list real-world validation only as future work.","Beyond the paper: the reward design is not ablated, so it is unknown whether the sparse foot-contact rewards or the dense distance and orientation terms do the heavy lifting; isolating each term would make the method more portable."],"forward_implications":["A complete autonomous skateboarding pipeline becomes possible by appending the learned mounting skill to an existing riding controller.","The same reverse-curriculum schedule could be applied to other contact-rich tasks with a known goal configuration, such as climbing onto a platform or entering a narrow bay.","Operators no longer need to position the robot precisely before the mounting attempt, since the trained policy tolerates variation in starting position and orientation.","Because the policy transfers from a fixed to a moving board, training effort can be concentrated on the contact-rich mounting phase rather than on modelling the board's rolling dynamics.","The observation of board-edge points at 10 cm intervals suggests the policy can accommodate skateboards of different shapes without changing the network."],"supporting_citations":[{"why":"This citation supplies the reverse-curriculum generation method the paper adapts to the mounting task.","marker":"[24]"},{"why":"This citation provides the simulation framework in which the quadruped and skateboard are trained.","marker":"[26]"},{"why":"This citation provides the base quadruped locomotion environment that the authors extend with a skateboard.","marker":"[25]"},{"why":"This citation demonstrates prior quadruped skateboarding with the robot already on the board, establishing the gap the paper targets.","marker":"[7]"},{"why":"This citation shows skateboarding for a bipedal robot that also begins on the board, reinforcing the claim that mounting is an unaddressed problem.","marker":"[5]"},{"why":"This citation presents an optimization-based skateboarding controller that also assumes the robot is already on the board, another baseline for the gap.","marker":"[8]"}],"fun_headline_variants":["Quadruped robot mounts skateboard via reverse curriculum","Robot dog learns skateboard mounting in sim, transfers to real","Reverse curriculum gets robot dog on the board","From rigid to moving board: robot dog mounts skateboard","Sim-to-real: quadruped mounts skateboard with reverse RL"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim assumes that a policy trained with the skateboard fixed in place generalizes to a skateboard that can roll and turn, even though the policy never experiences board motion for most of its training.","fun_headline_variants_meta":{"raw":{"variants":["Quadruped robot mounts skateboard via reverse curriculum","Robot dog learns skateboard mounting in sim, transfers to real","Reverse curriculum gets robot dog on the board","From rigid to moving board: robot dog mounts skateboard","Sim-to-real: quadruped mounts skateboard with reverse RL"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000419,"raw_usage":{"total_tokens":2101,"prompt_tokens":834,"completion_tokens":1267,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":450,"completion_tokens_details":{"reasoning_tokens":1188}},"tokens_in":450,"tokens_out":1267,"duration_ms":9434,"temperature":1.0,"reasoning_tokens":1188,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:38:44.604513+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained policy on a physical quadruped facing a real skateboard under the same starting offsets used in simulation; if the success rate is near zero because the board shifts during foot placement, the transfer claim fails. A cheaper in-simulation test is to vary skateboard mass, wheel friction, and bushing stiffness beyond the training distribution and watch for success collapse.","supporting_citations":[{"cited_title":"Re- verse curriculum generation for reinforcement learning,","cited_arxiv_id":null,"evidence_quote":"This citation supplies the reverse-curriculum generation method the paper adapts to the mounting task."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"This citation provides the base quadruped locomotion environment that the authors extend with a skateboard."},{"cited_title":"A bipedal walking robot that can fly, slackline, and skateboard,","cited_arxiv_id":null,"evidence_quote":"This citation shows skateboarding for a bipedal robot that also begins on the board, reinforcing the claim that mounting is an unaddressed problem."},{"cited_title":"Optimization based dynamic skateboarding of quadrupedal robot,","cited_arxiv_id":null,"evidence_quote":"This citation presents an optimization-based skateboarding controller that also assumes the robot is already on the board, another baseline for the gap."}],"review_version":1}