{"id":"b7f816b5-388e-4c3a-8900-bd55931109e8","arxiv_id":"2501.00507","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"DRGBT is a real-time sampling-based planner that uses an EDF scheduler and dynamic expanded bubbles to certify collision-free splines under bounded obstacle velocity, demonstrated on a 6-DoF arm with moving objects and humans.","lead":"This paper presents DRGBT, a sampling-based motion planner that runs in real time on a single CPU and slows or stops a six-joint robot arm before it hits moving obstacles, including people. It adds a scheduling analysis and a new safety structure called dynamic expanded bubbles, then validates the approach in simulation and on a real xArm6 arm.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The hard real-time schedulability claim is contradicted by the paper's own timing data: measured worst-case T1 ≈ 70–73 ms exceeds the chosen T* = 50 ms, so the EDF condition in Eqs. (2)/(4) is not met.","rationale":"I read the paper as making two headline claims: (i) a scheduling framework that guarantees hard real-time execution on a single CPU core, and (ii) Theorem 1 guaranteeing safe motion under bounded obstacle velocity. The reader's REJECT focuses on (i), and I agree that this is the most load-bearing concern because it is the paper's formal basis for hard real-time operation and it is invalid at the chosen operating point. Eq. (4) is derived from e2 = T - e1; if e1 > T, e2 is negative and the EDF utilization condition cannot hold. The reported e1 ≈ 70/73 ms (Fig. 11) exceeds T* = 50 ms chosen in Sec. V-D, and the paper itself concedes e1 grows with obstacle count. Allowing the scheduler to truncate T1 changes the algorithm and does not provide a fresh WCET bound; preemption of T1 can also weaken the safety argument because generateGBur and updateCurrState are inside T1. Therefore the claim of hard real-time operation at 20 Hz, and certainly at 100 Hz, is unsupported. I also note that Theorem 1's safety guarantee is stated without a proof and relies on discrete checking in Algs. 2 and 3 with time step Delta_t, which the paper acknowledges in Sec. IV-D; this is a second serious concern, but the schedulability contradiction is sufficient on its own and is the cleanest decisive objection. The experimental study is extensive and suggests the planner works well in practice; my objection is to the strength of the guarantees claimed, not to the quality of the empirical work.","tokens_in":32006,"tokens_out":5165,"duration_ms":54405,"concrete_test":"Instrument the released C++ implementation and run the same benchmark suite (15 scenario types x 1000 runs) with T* = 50 ms and u1 = 1, recording every T1 deadline miss; if any run exceeds 50 ms, or if the measured maximum e1 over all runs exceeds T*, then the EDF condition (4) is violated and the hard real-time claim is falsified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central 'hard real-time' claim rests on the EDF sufficient condition in Sec. III: sum_i e_i / min(D_i, T_i) <= 1. With D1 = D2 = T and e2 = T - e1, Eq. (4) reduces to e1/T + (T - e1)/T = 1, which is only valid if e1 <= T. The paper's own measurements in Sec. V-B and Fig. 11 give the maximum Task 1 execution time as e1 ≈ 70 ms for DRGBT and ≈ 73 ms for DRGBT-safe over 15,000 randomized runs, and the text explicitly concedes that with more than 50 obstacles, e1 'will likely become larger.' Yet Sec. V-D selects T* = 50 ms for DRGBT(1), i.e., 20 Hz operation, and Sec. VIII claims operation at frequencies up to 100 Hz (T = 10 ms). At these operating points e1 > T, so e2 = T - e1 is negative and the schedulability condition fails; Task 1 cannot be guaranteed to meet its deadline. The paper's response is to allow the scheduler to interrupt T1 (Sec. V-C/D), but then the empirical 'worst case' is no longer an upper bound for the constrained execution, and no new WCET is established. Moreover, the CDF evidence is probabilistic: generateGBur completes within 30 ms in only 99.9859% of cases, so a hard guarantee cannot be derived from these data. This is a direct internal inconsistency in the paper's primary real-time contribution.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents DRGBT, a sampling-based motion planner for robotic manipulators in dynamic environments, along with a scheduling framework claimed to enable hard real-time execution on a single CPU core, a new safety structure called dynamic expanded bubbles (DEBs), and both extensive simulations and real-robot experiments. The algorithm decomposes planning into two tasks (T1: local motion computation; T2: global replanning), uses EDF scheduling, and introduces a chain of DEBs to guarantee safe motion under bounded obstacle velocity. The paper reports a 15,000-run randomized simulation study, comparisons against RRTX and MARS, and physical experiments with a UFactory xArm6 manipulator and depth cameras.","tokens_in":32373,"tokens_out":2731,"duration_ms":30850,"significance":"If the hard real-time and safety claims were fully substantiated, the paper would be a useful contribution to real-time motion planning for manipulators in human-robot coexistence scenarios. The paper's strengths include a publicly available implementation, a large randomized evaluation over obstacle counts up to 50, a comparative study with state-of-the-art planners, and a real-robot validation with a sensor-based pipeline. The central advertised contribution, however, is the hard real-time property, and that claim is directly contradicted by the paper's own measurements. The safety theorem is also stated without a proof. These issues affect the paper's main thesis and cannot be resolved by minor editing alone.","major_comments":[{"comment":"The hard real-time schedulability claim is internally inconsistent with the reported timing data. Equation (4) requires e1 <= T for the EDF sufficient condition to be meaningful, since e2 = T - e1 becomes negative otherwise. However, Sec. V-B reports a maximum Task 1 execution time of e1 ≈ 70 ms for DRGBT and ≈ 73 ms for DRGBT-safe over 15,000 runs, and Sec. V-D selects T* = 50 ms (20 Hz), with Sec. VIII additionally claiming operation up to 100 Hz (T = 10 ms). At these operating points e1 > T, so T1 cannot be guaranteed to meet its deadline in all runs; the scheduler must interrupt T1, and no worst-case execution time for the partially executed task is established. The CDF evidence is probabilistic (e.g., generateGBur completes within 30 ms in 99.9859% of cases), not a hard bound. Furthermore, the text concedes that e1 'will likely become larger' with more than 50 obstacles, so the empirical maximum is not a valid WCET. This undermines the paper's primary real-time contribution.","section":"Sec. III, Eq. (4); Sec. V-B; Sec. V-D; Sec. VIII"},{"comment":"Theorem 1 is stated as a guarantee of safe motion but is not proved; the surrounding text only describes the algorithmic procedure (Algs. 2 and 3) and states that the spline is safe if it lies inside a chain of DEBs. The theorem also depends on a discrete time step Δt in the collision checking, while the guarantee is phrased as continuous-time safety. Moreover, the paper later allows type II collisions (zero-speed contacts), as stated in Sec. IV and Sec. V-D, so the claim 'guaranteed safe motion' is weaker than it appears: the robot may still collide once stopped. A precise statement with a proof or a precise characterization of the conditions under which type II collisions can occur is needed. Without this, the safety contribution is not established at the level claimed by Theorem 1.","section":"Sec. IV, Theorem 1"},{"comment":"The claim that the algorithm is capable of real-time operation at frequencies up to 100 Hz is not supported by the presented schedulability analysis. The simulation results in Sec. V-D show that DRGBT-safe performs best for T in the range 10–20 ms (i.e., 50–100 Hz), but these values are smaller than the measured maximum e1 ≈ 73 ms, so the EDF condition of Eq. (4) is violated. The 'sweet spot' for DRGBT(1) is reported at T* = 50 ms, which itself already violates the measured worst-case e1. Consequently, the paper does not demonstrate that any chosen operating point satisfies hard real-time constraints; at best it shows soft real-time behavior under the tested scenarios.","section":"Sec. V-D and Sec. VIII"}],"minor_comments":[{"comment":"The text says 'wmin = wmin = 0.5' where one occurrence should presumably be a different symbol or the sentence should be rephrased.","section":"Sec. V-A"},{"comment":"The symbol '≺' used for element-wise comparison is nonstandard; the manuscript would benefit from a formal definition or a sentence explaining the notation.","section":"Sec. IV, Definition IV.3"},{"comment":"The pseudocode in Alg. 1 has the input variable 'saf eon', which appears to be a spacing artifact; this should be corrected to 'safe_o n' or 'safeon' consistently.","section":"Sec. II-E"},{"comment":"The term 'hard real-time' is applied to T2 even though T2 is forcibly terminated at the next release of T1; this is an unconventional use of the term and should be clarified.","section":"Sec. III-A and Alg. 1"},{"comment":"The figures (Figs. 13 and 14) are informative but the captions do not clearly state the meaning of the 'criteria' axes or the distinction between solid and dotted lines; more detailed captions would help.","section":"Sec. V-D"}],"recommendation":"reject","confidential_remarks":"The paper has a solid empirical side and a working system, but the central hard real-time claim is not merely overenthusiastic: the paper's own data show that the chosen deadlines are missed in some runs, and the scheduling condition (4) is violated. The safety theorem is stated without proof, and the type-II-collision caveat further weakens the guarantee. These are load-bearing issues that affect the core advertised contributions. The authors could potentially reframe the contributions as 'soft real-time' or 'bounded-interruption' and provide a formal safety proof, but that would be a substantial revision of the paper's scope and claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know that the dynamic expanded bubble (DEB) is a genuine new idea, and the experimental effort is substantial—15,000 randomized runs, real robot tests with humans, and open-source code. The paper does what it says on the tin: it extends the authors' earlier DRGBT work with a safety structure that scales with obstacle velocity, and it shows the planner beating MARS and RRTX in head-to-head trials. That is real value, and the authors are transparent about their setup and limitations.\n\nThe soft spot is the central 'hard real-time' claim, and it is not a minor issue. The schedulability condition in Eq. (4) requires e1/T + e2/T <= 1, which only works if e1 <= T. The paper's own measurements give e1 ≈ 70–73 ms maximum, yet they select T* = 50 ms for the main operating point. That means Task 1 misses its deadline in some runs. Their mitigation is to let the scheduler interrupt generateGBur when time runs out, but then the measured e1 is no longer a worst-case bound for the actual constrained execution, and the CDF shows generateGBur completes within 30 ms only in 99.9859% of cases—a probabilistic statement, not a hard guarantee. So the \"hard\" part is not earned. This could be fixed by reframing as soft real-time or by providing a certified WCET through measurement-based analysis with appropriate margins, but as written, the claim is overstated.\n\nThe safety theorem is also stated without a proof and depends on discrete collision checking with a time step Δt. The paper acknowledges this and points to the capsule approximation as providing margin, which helps, but \"guaranteed safe motion\" is too strong; the guarantee is conditional on the discretization and on the obstacle velocity bound not being exceeded—and in Scenario 4 they report the human arm reached nearly 1.8 m/s against a bound of 0.5 m/s. Again, the idea is sound, the execution is honest, but the wording oversells.\n\nWho should read this? Anyone working on sampling-based planning for dynamic manipulator environments, especially on distance-aware safety structures. The DEB/DGBur machinery is worth studying and possibly extending. The paper deserves a serious referee: the flaws are in the claims, not in the core approach. I would send it to review, but with the expectation that the real-time section gets rewritten and the theorem either gets a real proof or gets softened to a proposition with clearly stated sufficient conditions.","headline":"Solid new safety structure and extensive experiments, but the hard real-time claim is contradicted by the paper's own timing data and needs reframing.","tokens_in":32953,"tokens_out":1530,"would_cite":true,"duration_ms":17161,"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":"The paper claims that a sampling-based manipulator planner can meet hard real-time deadlines on a single CPU core and, using dynamic expanded bubbles, guarantee safe motion whenever obstacle speeds are bounded.","keywords":["real-time motion planning","dynamic environments","sampling-based planning","dynamic expanded bubble","generalized bur","robotic manipulators","hard real-time scheduling","human-robot collaboration"],"falsifier":"Run the same timing protocol but with, say, 60 obstacles moving at 1.6 m/s and force replanning every iteration at $T = 50$ ms; if Task 1 ever exceeds its 50 ms budget, the hard real-time claim as stated is false for that operating point.","tokens_in":31784,"feed_emoji":"🤖","tokens_out":7762,"duration_ms":76164,"temperature":0.7,"pith_summary":"The paper is trying to establish that a sampling-based motion planner for multi-DoF robotic manipulators can run under hard real-time constraints on one CPU core and still carry a provable safety guarantee in environments where obstacles move unpredictably, as long as their speed is bounded. It upgrades the DRGBT planner with a two-task earliest-deadline-first schedule and with a new geometric structure, the dynamic expanded bubble, which shrinks the known-safe region in configuration space by the distance an obstacle could cover during the planning interval. If the central claims hold, a six- or ten-DoF arm equipped with two depth cameras can replan at 20 Hz and still guarantee that any collision happens only after the robot has stopped, with joint velocity, acceleration, and jerk limits respected. The paper supports these claims with a 15,000-run randomized simulation study, comparisons against established replanning algorithms, and real experiments including human presence in the workspace.","feed_headline":"Guaranteed safe robot motion at 20 Hz on one CPU","feed_subtitle":"Dynamic expanded bubbles turn obstacle speed into a safety margin, so the arm can stop before contact without heavy hardware.","key_machinery":"The central object is the dynamic expanded bubble (DEB), a configuration-space region defined around the current configuration by the inequality $\\sum_{j=1}^i r_{i,j}\\,|y_j - q_j| + v_{\\mathrm{obs}} t \\le d_i$ for every link $i$ and for time $t$ within the planning horizon. Here $d_i$ is the measured minimal distance from link $i$ to the obstacles, $r_{i,j}$ is a radius that bounds how far points on the robot can move when joint $j$ rotates, and $v_{\\mathrm{obs}}$ is the assumed maximum obstacle speed. The bubble is exactly the set of configurations the robot can reach during time $t$ without any link being able to collide, even if an obstacle rushes straight toward it at full speed; higher $v_{\\mathrm{obs}}$ or smaller $d_i$ shrinks the bubble. The algorithm chains these bubbles along the candidate spline by updating the separating planes between robot links and obstacles, producing a dynamic generalized bur and a sufficient condition for the whole composite trajectory — current spline plus emergency stop — to be collision-free. This condition, together with a quintic spline parameterization satisfying joint velocity, acceleration, and jerk limits, is what Theorem 1 turns into the guarantee of safe motion.","core_discovery":"The paper's central claim is that DRGBT, a sampling-based planner built on generalized burs of free configuration space, can be decomposed into two tasks — periodic computation of the next configuration and sporadic replanning — and scheduled by earliest-deadline-first on a single CPU core without GPUs or heavy parallelization. The second central claim is Theorem 1: if the current configuration is safe and every obstacle moves at speed at most $v_{\\mathrm{obs}}$, then the planner always produces either a safe spline to a new configuration that respects the kinematic constraints $\\mathcal{K}$, or an immediate emergency stop along a safe spline; a solution satisfying $\\mathcal{K}$ always exists inside the chain of connected dynamic expanded bubbles. In the safe variant, collisions can only occur after the robot has stopped, a property the paper calls type II collision. The authors report real-time operation at 20 Hz with frequencies up to 100 Hz in the simulation analysis, and they present randomized trials and real-robot scenarios supporting the claims.","pith_inferences":["A testable extension is to feed detected obstacle velocity directions into the bubble computation; the paper's equations show the safety margin could be relaxed by the cosine of the angle between the approach direction and the obstacle velocity, at the cost of monitoring the velocity vector's validity.","The two-task earliest-deadline-first scheduling idea transfers to other sampling-based planners: any planner whose per-iteration routines have an empirically bounded worst-case execution time could use the same logic, provided the measured bound is accepted as the true worst case.","Running the planner at 100 Hz while the camera supplies perception at 20 Hz means several iterations reuse stale obstacle data; combining the dynamic expanded bubble margin with constant-velocity prediction could make high-frequency replanning useful without weakening the safety claim."],"forward_implications":["At the chosen operating point, DRGBT(1) is claimed to meet all deadlines at an iteration time $T^* = 50$ ms, i.e., 20 Hz replanning, on a single CPU core; DRGBT-safe is reported to run at 50–100 Hz for small $T$.","With safety enabled, any collision that occurs is of type II, meaning contact only after the robot has already come to a stop, provided obstacle speeds stay below $v_{\\mathrm{obs}}$ and no obstacle appears outside the occupied halfspace region.","The robot automatically slows down as obstacles approach: smaller measured distances $d_i$ or higher assumed obstacle speed produce shorter safe splines and lower peak velocities.","Replanning can be interrupted or skipped and the robot can still reach the goal using only the local horizon, though performance degrades; replanning remains beneficial for global progress toward the goal.","In the paper's randomized comparison, DRGBT reports higher success rates than the MARS and RRTX baselines across obstacle counts and DoF values, with shorter algorithm times and comparable or shorter path lengths."],"supporting_citations":[{"why":"Defines the generalized bur of free C-space that DRGBT uses for fast local exploration and replanning.","marker":"[51]"},{"why":"Introduces expanded bubbles and per-link distance profiles, the foundation of the dynamic expanded bubble.","marker":"[57]"},{"why":"Establishes the original bubble of free C-space from workspace distance information.","marker":"[54]"},{"why":"Introduces burs of free C-space as collision-free spines, which the dynamic generalized bur extends.","marker":"[55]"},{"why":"Presents the original DRGBT algorithm and its horizon-and-spine machinery, extended here with scheduling and safety analysis.","marker":"[50]"},{"why":"Provides the RGBMT* replanner used in the DRGBT(1) variant and in the timing studies.","marker":"[56]"},{"why":"Gives the earliest-deadline-first schedulability condition used to prove the two-task hard real-time schedule.","marker":"[58]"},{"why":"Baseline RRTX algorithm compared in the randomized and scenario trials.","marker":"[30]"},{"why":"Baseline MARS replanning strategy compared in the randomized and scenario trials.","marker":"[46]"}],"fun_headline_variants":["DRGBT: Safe arm motion at 20 Hz on one CPU","Guaranteed safe motion in dynamic scenes, no GPU","Dynamic bubbles guarantee safe arm stops at 20 Hz","Real-time safe planning for arms without heavy hardware","Single-core planner guarantees safety in dynamic scenes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The hard real-time guarantee assumes that the timing measurements from the paper's 15,000 randomized trials are a valid worst-case bound on Task 1 for every future run, even though the paper notes that more than 50 obstacles will likely make the task slower, and the chosen operating period of 50 ms is smaller than the measured maximum of about 70 ms.","fun_headline_variants_meta":{"raw":{"variants":["DRGBT: Safe arm motion at 20 Hz on one CPU","Guaranteed safe motion in dynamic scenes, no GPU","Dynamic bubbles guarantee safe arm stops at 20 Hz","Real-time safe planning for arms without heavy hardware","Single-core planner guarantees safety in dynamic scenes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000614,"raw_usage":{"total_tokens":2841,"prompt_tokens":923,"completion_tokens":1918,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":539,"completion_tokens_details":{"reasoning_tokens":1841}},"tokens_in":539,"tokens_out":1918,"duration_ms":13316,"temperature":1.0,"reasoning_tokens":1841,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:49:24.782143+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same timing protocol but with, say, 60 obstacles moving at 1.6 m/s and force replanning every iteration at $T = 50$ ms; if Task 1 ever exceeds its 50 ms budget, the hard real-time claim as stated is false for that operating point.","supporting_citations":[{"cited_title":"Improved C-space exploration and path planning for robotic manipulators using distance information,","cited_arxiv_id":null,"evidence_quote":"Defines the generalized bur of free C-space that DRGBT uses for fast local exploration and replanning."},{"cited_title":"Path planning for robotic manipulators using expanded bubbles of free C-space,","cited_arxiv_id":null,"evidence_quote":"Introduces expanded bubbles and per-link distance profiles, the foundation of the dynamic expanded bubble."},{"cited_title":"Quinlan, Real-time modification of collision-free paths","cited_arxiv_id":null,"evidence_quote":"Establishes the original bubble of free C-space from workspace distance information."},{"cited_title":"Burs of free C- space: A novel structure for path planning,","cited_arxiv_id":null,"evidence_quote":"Introduces burs of free C-space as collision-free spines, which the dynamic generalized bur extends."},{"cited_title":"Path planning for robotic manipulators in dynamic environments using distance information,","cited_arxiv_id":null,"evidence_quote":"Presents the original DRGBT algorithm and its horizon-and-spine machinery, extended here with scheduling and safety analysis."},{"cited_title":"Asymptotically optimal path planning for robotic manipulators: Multi-directional, multi-tree approach,","cited_arxiv_id":null,"evidence_quote":"Provides the RGBMT* replanner used in the DRGBT(1) variant and in the timing studies."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the earliest-deadline-first schedulability condition used to prove the two-task hard real-time schedule."},{"cited_title":"RRT x: Asymptotically optimal single-query sampling-based motion planning with quick replanning,","cited_arxiv_id":null,"evidence_quote":"Baseline RRTX algorithm compared in the randomized and scenario trials."},{"cited_title":"Anytime informed multi-path replanning strategy for complex environments,","cited_arxiv_id":null,"evidence_quote":"Baseline MARS replanning strategy compared in the randomized and scenario trials."}],"review_version":1}