{"id":"d2d07a0b-1737-48a3-ab95-ff711c2874ad","arxiv_id":"2510.18085","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A team of robots can be trained by imitation from single-agent demonstrations collected round-robin, one robot at a time, matching or beating joint-action behavior cloning.","lead":"This paper introduces Round-Robin Behavior Cloning (R2BC), a way to teach a team of robots by having a human teleoperate one robot at a time while the others act on their current learned policies. It reports that R2BC matches or beats imitation learning trained on privileged synchronized demonstrations in simulation and on two physical robot tasks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Online co-adaptation lacks a convergence guarantee; the central claim rests on empirical plateaus that may not reflect a stable joint policy.","rationale":"The reader identified the same load-bearing concern: R2BC assumes the online co-adaptation loop converges, yet no convergence or stability guarantee is provided. This is the most critical assumption because the method's entire premise is that iterative single-agent demonstrations, with teammate policies as a moving target, yield a coherent multi-agent policy. The paper's own conclusion admits this gap. The simulation learning curves and ablations provide strong empirical support, but they cover only fixed, relatively small demonstration budgets; without longer-horizon evidence or a theoretical analysis, the claim that R2BC 'effectively trains' and 'reliably outperforms' is conditional on the loop not exhibiting pathological cycling in other regimes. I considered the physical-robot overclaim (navigation zero-shot p=0.077 yet conclusion says 'reliably outperforms') as an alternative concern, but it is secondary to the core algorithmic soundness issue. The reader's verdict of CONDITIONAL is appropriate; no verdict change is needed.","tokens_in":11473,"tokens_out":10123,"duration_ms":89853,"concrete_test":"Run R2BC on the Buzz Wire task (N=2, coupled dynamics) for 5x the current demonstration budget (e.g., 500 instead of 100) and record per-agent action distributions and team reward every 50 demonstrations. Compute the mean KL divergence between successive policy updates. If the KL divergence does not decay to near zero and/or evaluation reward oscillates rather than plateauing, the round-robin co-adaptation loop has not converged, undermining the claim that R2BC reliably trains a coherent team policy.","verdict_should_be":"UNCHANGED","load_bearing_attack":"R2BC's core mechanism (Algorithm 1) is an online, round-robin co-adaptation process: each agent's demonstrations are collected while the other agents execute their current learned policies, and policies are updated periodically. This creates a moving-target problem: after each update, the data previously collected for agent i was generated under a different joint policy, so the demonstration distribution for every agent shifts. The paper claims R2BC reduces covariate shift (Section V-C, Figure 3) and shows learning curves that increase and plateau (Figure 2), but there is no analysis of whether this alternating-update process converges to a fixed point, cycles, or diverges if run beyond the fixed demonstration budget. The conclusion explicitly concedes: 'developing theoretical guarantees for convergence and covariate shift reduction would further reinforce the intuition provided in this work.' Without such a guarantee, the central claim that R2BC 'enables a single human operator to effectively train multi-robot systems' is only supported for the specific, finite budgets tested. The concern is not merely theoretical: in tasks with coupled dynamics (e.g., Buzz Wire, where actions of one agent directly affect the other), the agents could in principle chase each other's changing policies indefinitely, producing oscillatory or unstable behavior that the current fixed-budget experiments would not reveal. The paper's empirical evidence is strong within the tested range, but it does not establish that the method reliably trains coherent team behavior beyond that range.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Round-Robin Behavior Cloning (R2BC), a multi-agent imitation learning method that trains a team of agents from single-agent demonstrations provided sequentially by one human. While one agent is teleoperated, the other agents act according to their current learned policies; demonstrations are stored in per-agent buffers and policies are updated periodically in a round-robin fashion. The method is evaluated on four VMAS simulated tasks (Navigation, Balance, Buzz Wire, Transport) against three oracle baselines: Joint Behavior Cloning (JBC), DAgger, and DART, all of which use privileged centralized demonstrations from a MAPPO expert. R2BC matches or exceeds these baselines. Ablations show that both decentralization and online learning are essential. The method is also deployed on physical HeRo+ robots in navigation and block-pushing tasks with real human demonstrations, where R2BC is reported to outperform centralized JBC by factors of 3.25x and 5.9x, respectively. The central claim is that R2BC provides a practical alternative to coordinated joint-action demonstrations for multi-agent IL.","tokens_in":11753,"tokens_out":6201,"duration_ms":58411,"significance":"The problem setting—multi-agent IL from single-agent demonstrations—is novel, well-motivated, and practically relevant. The paper provides a simple, effective algorithm and backs it with a solid empirical methodology: 10 seeds with 99% confidence intervals in simulation, multiple ablations, physical robot experiments with human demonstrations, and a public supplement with code and videos. The result that a single human can train a cooperative team without synchronous joint-action demonstrations is an important step toward real-world multi-robot teaching. The main weakness is the absence of a convergence or stability analysis for the online co-adaptation loop, which is explicitly acknowledged in the conclusion. The empirical evidence is strong within the tested finite demonstration budgets, but the generality of the method beyond these budgets remains unexamined. Overall, the paper is a meaningful contribution to multi-agent IL and human-robot interaction, provided the authors address the concerns below.","major_comments":[{"comment":"The core mechanism is an online alternating-update process: each agent's demonstrations are collected while teammates execute their current policies, and policies are updated after every k rounds. This is a moving-target problem, and the paper provides no analysis of whether the process converges, cycles, or diverges if run beyond the fixed demonstration budgets. The conclusion explicitly concedes the lack of theoretical guarantees. For the central claim that R2BC 'enables a single human operator to effectively train multi-robot systems,' this is a load-bearing gap: the reported success could, in principle, be an artifact of early stopping before the co-adaptation destabilizes, especially in coupled-dynamics tasks like Buzz Wire. I recommend either (a) adding experiments that run substantially longer than the current budgets (e.g., 2–4x more demonstrations) and report whether performance","section":"Algorithm 1, Section IV, and Section VIII"},{"comment":"The comparison to DAgger and DART, which the text uses to support the claim that R2BC achieves 'a similar reduction in compounding errors,' is not precisely specified. It is stated that these baselines use 'oracle centralized corrections generated from the same multi-agent RL policy,' but it is unclear whether (i) the policies are decentralized or centralized, (ii) the expert provides joint-action labels at every timestep for all agents, and (iii) the same demonstration budget and round-robin schedule are used. If the baselines receive privileged joint-action corrections at every step, they are strong oracles, but this should be stated explicitly. If they instead use single-agent corrections, the setup should be described. Without this detail, the reader cannot assess whether the parity result is due to R2BC's specific mechanism or to the online correction paradigm generally. Please prov","section":"Section V-C, Figure 2"},{"comment":"The abstract states that R2BC outperforms centralized BC by 3.25x (navigation) and 5.9x (block pushing). In the navigation task, the zero-intervention difference is reported as not statistically significant (p = 0.0773, Section VII-B), with significance only emerging after allowing 1–2 interventions. The 3.25x figure is therefore an overstatement when presented without qualification in the abstract. The authors do acknowledge the lack of significance in the body, which is good, but the abstract should either include the caveat, report the p-value, or drop the navigation multiplier and present the significant intervention results instead. This matters because the abstract is the primary summary of the paper's contributions.","section":"Abstract and Section VII-B, Figure 6"}],"minor_comments":[{"comment":"The y-axis is labeled 'Loss Gap (Log)'—please specify the base of the log and define whether the gap is absolute or relative. Also clarify the training set used for computing training loss in the online methods (e.g., which demonstration buffer).","section":"Figure 3"},{"comment":"The update condition 'if c mod k = 0' means that when k=1, updates occur after every round, but when k>1, the first update occurs after the first round because c=0. This is fine, but it would be clearer to state that c counts full round-robin passes, and that updates occur at the end of every k-th pass. Also clarify whether all agents are updated simultaneously using all data collected so far or only the latest k episodes.","section":"Algorithm 1, line 15"},{"comment":"The sentence 'the authors provided two sets of 240 demonstrations for each task' is ambiguous. Does this mean 240 demonstrations per method per task (480 total per task), or 240 total split between methods? Please clarify. Also note that the human demonstrators are authors; this is acceptable for a proof of concept, but a brief discussion of potential bias and a plan for a user study with naive subjects would strengthen the paper.","section":"Section VII-A"},{"comment":"The 'Offline R2BC (No-Op Agents)' variant assumes that non-demonstrating agents execute no action. In many multi-agent environments, a 'no-op' may not be a valid action or may have a different effect than standing still. Please specify how no-op actions were implemented and whether the environment supports them.","section":"Section VI, Figure 4"},{"comment":"The claim 'we are the first to propose and deploy a behavior cloning method for multi-agent systems that learns solely from online single-agent demonstrations' is strong. While the related work appears thorough, consider softening to 'to the best of our knowledge' in the abstract as well, and ensure that the related work section explicitly contrasts R2BC with Fleet-DAgger and other interactive IL methods, which the paper already does but could be expanded.","section":"Abstract and Introduction"},{"comment":"There are minor formatting issues, such as 'Round-RobinBehaviorCloning' in the introduction and inconsistent spacing in the figure captions. A careful proofread is recommended.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper makes a solid empirical contribution and the main idea is novel and clearly presented. However, I believe the three major comments—particularly the missing stability analysis and the ambiguous online baselines—need to be addressed before publication. The abstract's overstated navigation result should also be corrected. The lack of a convergence guarantee is not a fatal flaw, but the authors should either provide supporting long-horizon experiments or explicitly limit their claims to the tested regime. The manuscript is within scope for a robotics journal and, after revision, would be a valuable addition."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper's problem framing is new and the empirical case is mostly convincing. The method—round-robin single-agent demonstration with online decentralized BC—lets one human teach a robot team, and the comparisons against oracle joint-action BC on four VMAS tasks are well done. The ablations back up the two design choices. The two physical robot tasks with real human demos help, though the zero-shot navigation difference isn't statistically significant.\n\nThe soft spots are real but not fatal. The biggest is the conclusion's phrasing: 'reliably outperforms' is too strong when the no-intervention navigation comparison gives p=0.077. With one or two interventions the difference is significant, and the block-pushing result is significant at zero interventions, so the pattern is likely real, but the authors should hedge that one specific result.\n\nThe lack of convergence guarantees is honest—they list it as future work—and I don't think it's a blocker. R2BC is an online IL algorithm, and we don't usually demand formal convergence for DAgger-style methods either; the relevant question is whether the finite-budget experiments show stable improvement, and they do. The stress-test worry about potential cycling beyond the tested budget is speculative; the paper only claims success within the tested budgets. Still, a sentence noting that behavior outside those budgets is unexplored would be fair.\n\nThe paper also defers some hyperparameters and training details to a website. That's annoying for reproducibility, but not a technical flaw.\n\nNo circularity: the same MAPPO oracle labels all methods, and R2BC is compared to external baselines, not to a quantity it defines.\n\nBottom line: this is a solid empirical contribution to multi-agent IL. It introduces a practical problem setting that others will build on. The authors could tighten one overclaim and add a caveat about scope. I'd send it to a strong venue and hope the reviewers push for the softening rather than a rewrite.","headline":"R2BC is a genuinely new problem setup with solid empirical support, but the paper overstates one real-world result and needs a softer conclusion.","tokens_in":12248,"tokens_out":2288,"would_cite":true,"duration_ms":20343,"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":"Round-Robin Behavior Cloning (R2BC) lets one human teach a robot team by teleoperating one agent at a time, matching or beating behavior cloning from privileged synchronized demonstrations.","keywords":["multi-agent imitation learning","behavior cloning","single-agent demonstrations","round-robin training","multi-robot teleoperation","covariate shift","cooperative multi-robot systems","sim-to-real transfer"],"falsifier":"Track the per-agent train-test loss gap and the change in teammate policies across successive rounds on a fixed task. If the loss gap grows or oscillates without bound as demonstrations accumulate—or if the gap does not shrink relative to an offline variant—the online co-adaptation loop is not converging and R2BC's advantage over joint behavior cloning would not be explained by covariate-shift reduction.","tokens_in":11335,"feed_emoji":"🤖","tokens_out":10216,"duration_ms":74431,"temperature":0.7,"pith_summary":"Multi-agent imitation learning has usually assumed that a demonstrator can provide synchronized, near-optimal actions for every agent at once—something a single human cannot actually do. This paper introduces Round-Robin Behavior Cloning (R2BC), in which a human teleoperates one agent at a time while the other agents execute their current learned policies, cycling through the team. The paper shows that this online loop matches or exceeds the performance of an oracle behavior cloning method trained on privileged synchronized joint demonstrations across four simulated cooperative tasks, and that policies trained from real human demonstrations outperform centralized behavior cloning by 3.25x and 5.9x on two physical robot tasks. The authors argue the reason is that the imperfect actions of non-demonstrated agents inject state diversity into the collected data, reducing covariate shift in the same way explicit noise injection does in single-agent methods. The result matters because it replaces an infeasible demonstration protocol with one any single operator can execute.","feed_headline":"Round-robin teaching beats synchronized robot demos","feed_subtitle":"One operator, one robot at a time: team training beats synchronized demos on real hardware.","key_machinery":"Round-Robin Behavior Cloning (R2BC), a training loop that cycles which single agent the demonstrator controls while the rest of the team executes its current policies, then clones each agent's demonstrations into a separate policy trained only on that agent's buffer. The load-bearing mechanism is teammate-induced state perturbation: because the non-demonstrated agents act suboptimally and are themselves being updated, the demonstrated agent sees a diverse stream of teammate behavior, which teaches corrective actions and reduces covariate shift the way explicit noise injection does in single-agent online imitation learning.","core_discovery":"R2BC's central claim is that a single human can train a coherent multi-agent policy through sequential single-agent demonstrations, provided the demonstrations are collected online in a round-robin loop: agent i is teleoperated while the other agents run their current cloned policies; the resulting observation-action pairs go into agent i's own buffer; and all policies are re-trained at regular intervals. Because the demonstrating agent is repeatedly exposed to states created by imperfect teammates, it collects corrective examples from a distribution close to what the team will actually experience at execution time. The paper argues this online co-adaptation, together with full decentralizat","pith_inferences":["The mechanism suggests a broader principle: a learning team can serve as its own data generator, with teammates' imperfections providing the corrective signal a teacher needs—so the teacher's update rate only needs to outpace teammate deterioration, a rate condition the paper does not analyze.","A testable extension would sweep the update frequency and round-robin order; if teammate-induced state diversity is the source of the benefit, the method should be robust to both, and the optimal schedule could guide deployment on larger teams.","The physical-robot demonstrations were collected by the paper's authors; an independent user study with non-expert teachers would test whether the 3.25x and 5.9x margins persist—a step the paper lists as future work."],"forward_implications":["A single human operator can train a multi-robot team with ordinary single-agent teleoperation; synchronized joint-action demonstrations are no longer a prerequisite for multi-agent imitation learning.","R2BC matches or surpasses behavior cloning from privileged joint demonstrations on four simulated cooperative tasks, so realistic demo collection does not have to cost task performance.","Policies trained from real human demonstrations and deployed on physical robots outperform centralized behavior cloning by 3.25x on navigation and 5.9x on block pushing, and short human interventions amplify the gap.","Ablations show that both decentralization and the online update loop are necessary—centralizing the policy or freezing teammate policies during collection degrades results—so the claimed mechanism is the online co-adaptation, not simply round-robin data collection."],"fun_headline_variants":["Round-robin demos let one human train multi-robot teams","Single operator, round-robin demos, multi-robot imitation","R2BC: One human trains a robot team via round-robin","Round-robin imitation: single human, multi-robot teams","One operator, sequential demos, multi-agent robot control"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The round-robin online loop is assumed to converge: as each agent's policy is updated, the non-demonstrated agents become a moving target, and demonstrations collected against them still support a coherent team policy—but the paper provides no convergence or stability guarantee.","fun_headline_variants_meta":{"raw":{"variants":["Round-robin demos let one human train multi-robot teams","Single operator, round-robin demos, multi-robot imitation","R2BC: One human trains a robot team via round-robin","Round-robin imitation: single human, multi-robot teams","One operator, sequential demos, multi-agent robot control"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000813,"raw_usage":{"total_tokens":3369,"prompt_tokens":681,"completion_tokens":2688,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":425,"completion_tokens_details":{"reasoning_tokens":2599}},"tokens_in":425,"tokens_out":2688,"duration_ms":17431,"temperature":1.0,"reasoning_tokens":2599,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T08:52:27.209444+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Track the per-agent train-test loss gap and the change in teammate policies across successive rounds on a fixed task. If the loss gap grows or oscillates without bound as demonstrations accumulate—or if the gap does not shrink relative to an offline variant—the online co-adaptation loop is not converging and R2BC's advantage over joint behavior cloning would not be explained by covariate-shift reduction.","supporting_citations":[],"review_version":1}