{"id":"3faaee20-1af3-485f-9694-b913e1cd9b2a","arxiv_id":"2411.11510","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A hierarchical controller simulates sequences of closed-loop behaviors in a physics engine and chooses a plan; the robot reached its target in 10 of 10 overtaking runs.","lead":"A University of Glasgow team built a robot planner that uses a physics simulator as 'innate knowledge' to test sequences of simple closed-loop behaviors before moving, then executes the best plan. In a proof-of-concept overtaking scenario, the planning robot reached its target in all 10 runs, while a reactive controller never did.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central feasibility claim rests on unvalidated Box2D simulation fidelity; the reported 'brushed the obstacle twice' execution error is direct evidence of a sim-to-real gap that Algorithm 1 cannot detect.","rationale":"The reader's weakest assumption is that the Box2D simulation, built from resampled ego-centric LiDAR points, correctly predicts real-world task outcomes. This is indeed the most load-bearing assumption: the entire planning mechanism in Algorithm 1 expands successor states based on simulated collisions and terminations, and the plan extracted by the guard in Section 2.2.4 is only as good as those simulations. The paper's own results provide a concrete red flag: the robot brushed the obstacle twice in the planning condition due to execution error accumulation, meaning the real trajectory deviated from the simulated one. I agree with the reader's assessment and see no additional concern that would overturn the CONDITIONAL verdict. The framework is a plausible proof-of-concept for a single scenario, but the simulation-to-real gap is neither quantified nor mitigated, so the central claim of reliable multi-step planning via physics-engine simulation is not yet established. The proposed test directly checks whether the simulation reproduces real outcomes; if it does, the reader's conditional concerns are substantially addressed. I therefore recommend keeping the verdict unchanged: conditional acceptance pending validation of simulation fidelity, stronger comparison with baselines, and release of artifacts.","tokens_in":4469,"tokens_out":3230,"duration_ms":36838,"concrete_test":"Re-run the overtaking experiment with an external motion-capture system and full LiDAR logging. For each of the 10 planning runs, reconstruct the Box2D scene twice: once from the initial resampled LiDAR exactly as described in Section 2.2.1, and once from the motion-capture ground-truth obstacle geometry. Compare, at each task transition, the simulated state sequence and collision flags against the real executed sequence. If the as-described simulation cannot reproduce the real brushing/collision events or the branch choices, the simulation-fidelity assumption fails and the central claim is not supported beyond the tested scenario.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central assertion is that the Configurator can plan by simulating sequences of closed-loop tasks in Box2D. For that claim to hold, the simulation must faithfully predict which task transitions are collision-free and which are interrupted by a disturbance. That assumption is introduced in Section 2.2.1, where objects are built from resampled ego-centric LiDAR points 'in the way of the present task'. No calibration against real sensor noise, no validation of simulated trajectories, and no quantification of geometric fidelity are reported. More importantly, the one real-robot experiment already shows a divergence: in Section 3 the authors report that 'the robot brushed the obstacle twice in the planning condition, both times in one particular scenario due to execution error accumulation'. This is exactly a case where the simulated outcome (a safe plan) differs from real execution (a collision/brushing event), yet the Configurator has no mechanism to notice or repair the discrepancy. Because Algorithm 1 expands states based on simulated collision and termination outcomes, a single mismatch can invalidate the extracted plan. The 10-run success rate only demonstrates that the simulated branches were correct in one specific overtaking geometry; it does not establish the general reliability of the physics engine as 'core knowledge'.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hierarchical robot planning framework in which low-level closed-loop behaviours ('tasks') are temporary sensor-motor loops, and a high-level 'Configurator' selects and sequences these tasks by simulating candidate sequences in a physics engine (Box2D), treated as 'core knowledge'. The Configurator constructs a cognitive map of states and transitions, assigns costs, and extracts a plan as a sequence of tasks. A proof-of-concept is implemented on a real robot in an overtaking scenario: in 10 runs the robot formulated a plan to reach a target 1 m away, with mean planning time 0.064 ± 0.009 s over a state space of about 33 states; two brushing events occurred during execution. A comparison with a single reactive controller shows the reactive agent never reached the target and collided 8 times, but the difference in collisions was not statistically significant (p = 0.111).","tokens_in":4678,"tokens_out":3695,"duration_ms":40425,"significance":"If the claims hold, the paper makes a modest conceptual contribution: using a physics engine as an internal model for closed-loop task sequencing, rather than as a purely geometric path planner, and demonstrating real-time planning in a simple scenario. The paper is commendably honest about its limitations, including the non-significant comparison and the execution brushing events. However, the empirical evidence is thin: a single scenario, a small number of runs, no detailed simulation-to-real validation, and no code or data release. The central feasibility claim—that a physics engine can serve as a reliable predictor for task-sequence outcomes—rests on an unvalidated assumption about simulation fidelity. The work is best viewed as a proof-of-concept that could be strengthened with more systematic validation.","major_comments":[{"comment":"The central claim that Box2D can serve as 'core knowledge' requires that the simulation faithfully predicts which task transitions are collision-free and which are interrupted. The paper gives no validation of this fidelity: no comparison of simulated trajectories against real ones, no account of how sensor noise is modelled, and no quantitative measure of geometric accuracy of the LiDAR-derived objects. The two brushing events reported in Section 3 are direct evidence of a sim-to-real gap: the plan selected from the simulated search led to an actual collision, yet Algorithm 1 has no mechanism to detect or correct this discrepancy because its state expansion relies on simulated collision outcomes. The 10/10 plan-formation rate only shows that the simulated branches were correct in this one geometry; it does not establish the reliability of the physics engine as a general predictive model. Please provide a validation study (e.g., comparing simulated and real task outcomes across multiple geometries and disturbances) and discuss how the Configurator could handle residual sim-to-real uncertainty.","section":"Section 2.2.1 and Algorithm 1"},{"comment":"The comparative evaluation against the reactive controller is underpowered: the paired T-value of -1.747 with p = 0.111 is not statistically significant, and the paper's wording correctly says the difference 'approached, but did not reach' significance. The qualitative statement that the reactive condition 'never reached the target' is descriptive and not backed by a rigorous test. For a proof-of-concept, this comparison is not load-bearing for the feasibility claim, but it also does not support a claim of superiority. The paper should either present the comparison as purely illustrative, or provide more runs, a matched reactive controller, and a significance test that addresses the binary 'reached target' outcome. In addition, please state explicitly how many of the 10 planning runs resulted in the robot actually reaching the target, since the current text only says a plan was formulated.","section":"Section 3"},{"comment":"Several formal details needed for reproducibility are missing or ambiguous. The reset R(q1, q2) is defined only for DN = H, yet the cost function φ(q) uses the distance to DN 'if present', implying that some expanded states have a non-empty DN. It is unclear whether such interrupted states are inserted into the cognitive map and, if so, whether they can be expanded (the text later says transitions from states with DN ≠ H are not permitted, but the state-space figure appears to include starred states). The termination condition in Algorithm 1 ('until no obstacle-free transitions exist out of qE or DG has been reached') is also under-specified: how is 'obstacle-free' determined from a simulation run, and what happens if the goal is unreachable? Clarify these points and ensure the formal definitions match the algorithm's behaviour.","section":"Definitions 1–4 and Algorithm 1"}],"minor_comments":[{"comment":"The text refers to states q6, q8, q10 and q20, but the figure labels are difficult to read and some labels cited in the text are not clearly visible in the printed figure. Please improve the figure legibility and ensure all referenced states appear.","section":"Section 2.1.2, Figure 2"},{"comment":"The cost function φ(q) = γ(q) + χ(q) is said to be 'normalised', but the normalization procedure is not stated. Specify the range of distances used (e.g., maximum distance in the current map) and how γ is computed when no interrupting disturbance DN occurs.","section":"Section 2.2.3"},{"comment":"Please provide details of the reactive controller used as a baseline (e.g., its sensor processing, control gains, and termination conditions) so that the comparison is meaningful. The current description does not allow a reader to judge whether the controller was reasonably tuned.","section":"Section 3"},{"comment":"The manuscript does not indicate whether code or experimental data will be made available. For a proof-of-concept that depends on a specific simulation pipeline, releasing the code and raw logs would substantially improve reproducibility and allow independent verification.","section":"General"},{"comment":"The term 'innate physics knowledge' is used both in the abstract and in the text. Since the physics engine is designed and programmed by the authors, 'a-priori physics model' or 'internal physics model' would be less overclaiming and better aligned with the actual contribution.","section":"Introduction and Conclusion"}],"recommendation":"major_revision","confidential_remarks":"The paper is a short proof-of-concept. The biggest risk is that the simulation-to-real fidelity is asserted rather than demonstrated, and the only available evidence of a gap (the brushing events) is acknowledged but not addressed. The non-significant statistical comparison weakens the evaluation but does not, by itself, invalidate the feasibility claim. I recommend major revision, not rejection, because the framework is interesting and the missing validation appears obtainable within the scope of a revised manuscript. The journal should also consider whether a single-scenario study with n=10 provides enough support for the claimed generality of the 'core knowledge' concept."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a modest but genuine proof-of-concept. The Configurator—a supervisor that uses Box2D to simulate sequences of temporary closed-loop tasks and then picks a plan—is a new architectural integration, and it works on a real robot in an overtaking scenario. That is worth a fair read.\n\nWhat the paper does well: the formalism is simple and readable; the implementation is concrete; the real-robot results are honestly reported, including the two brushing events and the non-significant comparison with the reactive baseline. Planning time of 0.064 s over a 33-state search is a nice concrete number. The idea of using simulation as 'core knowledge' to build a cognitive map without physical exploration is clearly stated.\n\nThe soft spots are in proportion. The load-bearing assumption is that the Box2D simulation, built from resampled LiDAR points, faithfully predicts which task sequences are safe. That is not validated beyond the reported geometry. The two brushing events are exactly the kind of sim-to-real mismatch that Algorithm 1 cannot detect, because it only checks simulated collisions and terminations. That said, the robot still reached the target in all ten runs, so for this narrow scenario the simulation was good enough. The paper does not claim generality, and the conclusion is appropriately modest.\n\nThe comparison with the reactive controller is weak: a single closed-loop behaviour is not a strong baseline, and the p-value of 0.111 means the paper cannot claim an advantage. That is a minor issue for a proof-of-concept, but a serious referee should ask for code and data release and for a validation of simulation fidelity against real trajectories.\n\nBottom line: this is a solid proof-of-concept, not a paradigm shift. It deserves peer review, but the revision should be expected to add artifacts and a stronger empirical section. I would give it a conditional accept at a good robotics conference, not a desk reject.","headline":"A clean proof-of-concept for simulation-based hierarchical planning; the sim-to-real gap is real but not fatal, and the paper deserves a serious referee.","tokens_in":5175,"tokens_out":2565,"would_cite":false,"duration_ms":24445,"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 physics engine can serve as a robot's built-in 'core knowledge' for multi-step planning.","keywords":["closed-loop behaviour","multi-step planning","core knowledge","physics simulation","cognitive map","best-first search","mobile robot overtaking","supervisory control"],"falsifier":"Run the same overtaking scenario with the obstacle placed at a series of offsets; for each task branch, compare whether the physics simulation predicts collision or safe termination with what the real robot does when executing that branch. A systematic divergence — for example, a simulated left-turn branch marked safe that hits the obstacle on the robot, or a simulated collision that never materialises — would falsify the core-knowledge assumption and make the extracted plan unreliable.","tokens_in":4273,"feed_emoji":"🤖","tokens_out":11911,"duration_ms":102411,"temperature":0.7,"pith_summary":"This paper tries to establish that robot planning can be treated as an input control problem solved in real time by combining temporary closed-loop behaviours ('tasks') with a physics simulation used as built-in 'core knowledge.' A supervising Configurator simulates candidate sequences of tasks, marks branches that collide, organises the safe ones into a searchable cognitive map, and extracts a plan as the lowest-cost adjacent path. In the overtaking demonstration the robot produced a plan in all 10 runs, with mean planning time $0.064 \\pm 0.009$ s over about 33 states, while a reactive single-loop controller never reached the target. The point is that the physics engine acts as the agent's own internal reasoning tool, so planning becomes simulated exploration rather than learned control.","feed_headline":"Robot plans an overtake in 64 ms via simulated physics","feed_subtitle":"All ten runs produced a plan; a single-loop controller never reached the target.","key_machinery":"The central object is the cognitive map $G=(Q,\\to)$ built from states $q=(T,D)$, where $T$ is a temporary closed-loop task and $D$ is a set of disturbances. The Configurator grows this map by best-first expansion using the cost function $\\varphi(q)=\\gamma(q)+\\chi(q)$; $\\gamma$ is the normalised distance from the start of the task to an interrupting disturbance $D_N$, and $\\chi$ is the normalised negative distance from the end of the task to the goal $D_G$. A guard $\\Psi$ assigns a supervisory control pattern to each transition, and a reset $R$ injects the goal disturbance into the next task unless the previous task was interrupted. The physics engine (Box2D, a 2D rigid-body physics engine), fed by resampled ego-centric LiDAR points and a robot model, supplies the simulated outcomes — collision or termination — for each frontier transition, and these simulated outcomes drive which states enter the map.","core_discovery":"The paper's central claim is that a hierarchical controller called the Configurator can plan multi-step ahead by using a physics engine as 'core knowledge' to simulate what would happen if each temporary closed-loop task were executed. Each task is a behaviour contingent on a disturbance; the Configurator injects the goal as a disturbance, simulates frontier states depth-first until a straight-drive state or an interrupting collision, and records results as states $q=(T,D)$ in a cognitive map. The plan is the adjacent sequence of states ending at the state with least cost $\\varphi(q)=\\gamma(q)+\\chi(q)$, where $\\gamma$ is normalised past distance to an interrupting disturbance and $\\chi$ is normalised negative distance to the goal. In the overtaking test the robot formulated a plan in all 10 runs; planning took $0.064 \\pm 0.009$ seconds across a state-space of $32.9 \\pm 1.758$ states; the reactive single-loop condition never reached the target and collided 8 times over 10 runs (collision-count difference $T = -1.747$, $p = 0.111$).","pith_inferences":["The paper demonstrates only one task library (straight and turn behaviours) and one overtaking scenario; a natural extension would vary obstacle layouts and task sets to see whether planning time and success rate degrade gracefully as the frontier grows.","An implicit consequence is that the whole framework inherits the fidelity of the physics simulation: if the simulated collision and termination predictions mismatch the real robot, the selected plan can fail, so the approach rewards investing in better robot and sensor models.","One could test whether the planning horizon extends to longer sequences by widening the task library or adding goal-directed heuristics, trading computation for plan quality.","The phrase 'innate physics knowledge' suggests a stronger claim than the experiments support: that a physics-engine simulation can substitute for learned world models in embodied agents. That substitution is an interpretation beyond the paper's overtaking proof-of-concept."],"forward_implications":["Multi-step planning becomes a matter of choosing among a library of closed-loop behaviours, avoiding both reward-based training and external trajectory tracking.","Planning is fast enough for real-time use: in the overtaking runs the state-space of about 33 states was explored and a plan extracted in $0.064 \\pm 0.009$ seconds.","The reset rule ties every newly created task to the goal disturbance, so the extracted plan is goal-directed unless a task is interrupted by an obstacle.","The physics engine supplies all branch outcomes during planning, so the robot does not need to physically explore the environment in order to build the cognitive map."],"supporting_citations":[{"why":"Supplies the model of a task as a hard-wired closed-loop sensor-effector controller.","marker":"[2]"},{"why":"Defines 'core knowledge' as innate physics and causality understanding, which the paper adopts as the rationale for a physics engine.","marker":"[4]"},{"why":"Provides the supervisory control formalism used by the Configurator's guard $\\Psi$.","marker":"[5]"},{"why":"Supplies the best-first cost function $\\varphi=\\gamma+\\chi$ and priority-queue expansion used to grow the cognitive map.","marker":"[6]"}],"fun_headline_variants":["Physics engine guides hierarchical robot planning in 64 ms","Hierarchical planner uses physics simulation for fast robot plans","Robot overtaking plan in 64 ms via simulated physics","Physics-based hierarchy plans overtake in milliseconds","Configurator plans overtake with simulated physics in 64 ms"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the physics simulation built from the robot's resampled LiDAR points and robot model predicts real collisions and task terminations accurately enough that a branch simulated as safe will also be safe on the real robot.","fun_headline_variants_meta":{"raw":{"variants":["Physics engine guides hierarchical robot planning in 64 ms","Hierarchical planner uses physics simulation for fast robot plans","Robot overtaking plan in 64 ms via simulated physics","Physics-based hierarchy plans overtake in milliseconds","Configurator plans overtake with simulated physics in 64 ms"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000469,"raw_usage":{"total_tokens":2295,"prompt_tokens":865,"completion_tokens":1430,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":481,"completion_tokens_details":{"reasoning_tokens":1352}},"tokens_in":481,"tokens_out":1430,"duration_ms":11639,"temperature":1.0,"reasoning_tokens":1352,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T18:25:08.879206+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same overtaking scenario with the obstacle placed at a series of offsets; for each task branch, compare whether the physics simulation predicts collision or safe termination with what the real robot does when executing that branch. A systematic divergence — for example, a simulated left-turn branch marked safe that hits the obstacle on the robot, or a simulated collision that never materialises — would falsify the core-knowledge assumption and make the extracted plan unreliable.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines 'core knowledge' as innate physics and causality understanding, which the paper adopts as the rationale for a physics engine."},{"cited_title":"Infrared Telescope Facility (IRTF) spectral library II: New indices in Y, J, H, and L atmospheric windows","cited_arxiv_id":"2007.00569","evidence_quote":"Provides the supervisory control formalism used by the Configurator's guard $\\Psi$."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the best-first cost function $\\varphi=\\gamma+\\chi$ and priority-queue expansion used to grow the cognitive map."}],"review_version":1}