{"id":"9e1022fc-3903-4159-ad37-f5b2065e7008","arxiv_id":"2501.01806","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"TRG-planner builds a direction-aware terrain risk graph and uses a risk-and-distance cost to plan safe, short paths for legged robots in unstructured environments.","lead":"This paper introduces TRG-planner, a path planner that builds a graph of safe, reachable terrain spots for legged robots and weights the paths between them by slope-based risk. It reports faster and safer navigation than standard planners in simulation and real-world quadruped tests, including a win at the ICRA 2023 Quadruped Robot Challenge.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (4) leaves PCA eigenvector orientation unspecified, so the sign of Rxi_lon/Rxi_lat can flip arbitrarily, making edge costs in Eq. (7) ill-defined and the claimed direction-aware risk non-reproducible.","rationale":"The reader's weakest_assumption identifies the single most load-bearing gap in the paper: Eq. (4) defines the traversal risk weight using PCA eigenvectors whose sign is arbitrary unless an orientation rule is specified. This is not a minor implementation detail; it directly affects the edge costs in Eq. (7), the admissibility of the heuristic, and the claimed 'direction-aware' risk that is the core novelty of TRG. Without knowing whether eξ_lon points from v_i to v_j or the reverse, the planner's behavior is not well-defined, and the empirical results, including the ICRA win, cannot be reproduced from the text alone. The concern is addressable—the authors could specify a convention or release code that fixes the orientation—so it does not invalidate the central idea, but it does warrant conditional acceptance with a request for clarification. The empirical evidence in Tables II-IV and the real-world demonstration provide partial support but do not resolve this formal gap. Therefore, the CONDITIONAL verdict remains appropriate, and no change to the reader's verdict is needed.","tokens_in":12664,"tokens_out":7522,"duration_ms":77406,"concrete_test":"Implement the risk computation of Eq. (4) on a synthetic 20-degree slope patch with nodes at the bottom and top. Compute PCA eigenvectors using a standard library (e.g., NumPy's eigh) without any sign alignment, then repeat after flipping the signs of the in-plane eigenvectors. If the resulting edge weight changes sign or magnitude such that the optimal path under Eq. (7) changes, the specification is incomplete. The authors should then state an orientation rule (e.g., eξ_lon aligned with p_j - p_i, and eξ_lat chosen to make a right-handed frame with the upward normal) and re-run the key experiments to confirm the reported values.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III.A.2 defines the risk weight w_ij in Eq. (4) as a linear combination of Rξ_lon and Rξ_lat, each computed as the negative inner product of a PCA eigenvector with the gravity vector g=[0,0,-1]. PCA eigenvectors are determined only up to sign, and the paper never states that eξ_lon is oriented along the travel direction from v_i to v_j or that its z-component is aligned with the uphill direction. Consequently, the sign of w_ij is not fixed by the definitions. Under a literal reading, a standard PCA routine can return eξ_lon pointing either uphill or downhill, changing w_ij from positive to negative. The cost function in Eq. (7) uses (Γw+1); if w < -1/Γ, the edge cost becomes negative, breaking the admissibility of the A* heuristic J and potentially causing the planner to prefer unrealistic paths. Moreover, the paper describes TRG as an undirected graph while also claiming direction-aware risk; without an explicit orientation rule, these two statements are inconsistent. This is the central definition on which the safety claims rest, so the planner is not fully specified as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents TRG-planner, a global path-planning method for unstructured terrain. The environment is represented as a Traversal Risk Graph whose nodes encode local standability/reachability and whose edges carry a traversal-risk weight derived from PCA of the local height map. The graph is built by wavefront sampling and updated hierarchically, and planning is performed by an A*-style search over a cost that combines Euclidean distance with the risk weight. The authors claim, based on simulation against A*, PRM*, and T-Hybrid and on real-world quadruped experiments (including the ICRA 2023 Quadruped Robot Challenge), that TRG-planner gives higher travel success, lower path risk, shorter paths, and much faster planning.","tokens_in":12885,"tokens_out":3031,"duration_ms":33694,"significance":"If the definitions were fully specified, the contribution would be of practical interest to field robotics: a real-time risk-aware global planner validated on a legged platform and in a competitive setting. The hierarchical graph construction and the wavefront propagation idea are reasonable engineering contributions, and the reported planning times are notably low. However, the central risk definition in Eq. (4) has a sign ambiguity that makes the planner's objective not fully defined as written, and the main safety metric in Eq. (8) is computed from the very edge weights that the planner minimizes. These issues must be resolved before the safety and efficiency claims can be assessed. The paper is currently not reproducible from the text alone.","major_comments":[{"comment":"The sign of the risk weight is not defined. The quantities Rξ_lon and Rξ_lat are defined as negative inner products of PCA eigenvectors with g = [0,0,-1]. PCA eigenvectors are determined only up to a sign, and the text does not specify any orientation rule relating êξ_lon to the travel direction from vi to vj or to the uphill direction. Under a literal reading, a standard PCA routine may return êξ_lon pointing in either direction, so w_ij can flip sign, and the graph can have negative or inconsistent edge weights. This is not a cosmetic issue: the cost update in Eq. (7) uses (Γw_{i+1,i} + 1), which becomes negative for w < -1/Γ, breaking the admissibility of the heuristic J and potentially causing the planner to prefer unrealistic paths. The description of TRG as an undirected graph also conflicts with the claimed direction-aware risk, since a single scalar w_ij cannot represent different risk for traversal from vi to vj versus from vj to vi. The authors must specify an orientation rule for the eigenvectors, clarify whether risk is symmetric or directed, and show that the resulting edge weights are nonnegative.","section":"Section III.A.2, Eq. (4)"},{"comment":"The normalized path risk W is not an independent safety measure for the proposed planner. W is computed as the average of the TRG edge weights w_{i+1,i} along the planned path, and those same weights are exactly the quantities minimized by the TRG-planner cost function in Eq. (7). Consequently, a comparison of W between TRG-planner and the baselines is biased in favor of TRG-planner by construction, especially because the baselines never consider w in their objectives. The travel success rate Strav and the deviation metric T are more independent, but W is presented as a main safety metric in Tables II and IV and in Section V. The authors should replace W with an externally defined terrain-risk measure (e.g., slope/roughness statistics along the traversed trajectory, or measured body orientation/contact events) or at least report the raw terrain properties of the actually traveled paths.","section":"Section IV.B, Eq. (8)"},{"comment":"The two risk-related parameters γ and Γ are tuned on the same test scenarios used for the final comparison, and the balanced strategy Γ=3.0 is selected post hoc from three tested values (1.0, 3.0, 10.0). This is a form of fitting to the test set, and it weakens the generality of the claim that the balanced strategy is best. In addition, Tables II-IV report only point estimates for metrics such as Spath, Strav, Lpath, W, and T, even though each scenario uses 100 randomly generated start-goal pairs; no error bars, confidence intervals, or statistical significance tests are provided. The authors should report variances or confidence intervals and should validate the parameter choice on a held-out set or via sensitivity analysis.","section":"Section V.A, Table II"},{"comment":"The graph construction and management rely on several informal terms that should be made precise for reproducibility: the 'vicinity' of the reference node and the radius rexp are clear enough, but the 'area covered by G' (QG in Eq. (5)) is not formally defined, and the frontier condition ui ∉ QG depends on this undefined set. Also, the statement that nodes are sampled 'following a uniform distribution on a circle' with radius rexp means the nodes lie on a ring rather than in a disk; this should be stated explicitly, since it affects graph connectivity and coverage.","section":"Section III.B and III.C"}],"minor_comments":[{"comment":"The definition T = 0 when negative is ad hoc; the authors should explain why a negative T arises and why it is truncated rather than treated as a meaningful value.","section":"Section IV.B"},{"comment":"The comparison with PRM* uses 'the same number of samples as the TRG nodes,' but the initialization time and roadmap quality of PRM* depend strongly on the sampling strategy and connection radius; the comparison would be clearer if the PRM* parameters were reported and justified.","section":"Section V.B, Table III"},{"comment":"In Eq. (4), the notation Rξ_dir is defined but not used afterward; the definitions of Rξ_lon and Rξ_lat would be easier to follow if the authors explicitly wrote out the components of êξ_lon and êξ_lat and the sign convention for uphill versus downhill.","section":"Section III.A.2"},{"comment":"The simulation environment is described as '50 m × 50 m × 6.9 m,' but Fig. 6(b) labels show '6.9 0.0' and the environment appears to be a height map; please clarify whether 6.9 m is the elevation range or the vertical extent of the map.","section":"Section IV.A.1"},{"comment":"The related work on traversability-aware planning is discussed adequately, but the comparison with the STEP planner [22] and SMUG planner [23] is only brief; since those methods also address risk-aware global planning, a paragraph contrasting their risk formulations with the TRG formulation would help.","section":"Section II.B"}],"recommendation":"major_revision","confidential_remarks":"The paper has a strong practical component, including a competition win, but the sign ambiguity in Eq. (4) is a load-bearing reproducibility issue that must be fixed before the work can be evaluated. The authors should also address the circularity of the W metric and the lack of variance reporting. I do not see evidence of intentional misconduct; the issues appear to be under-specification and evaluation-design choices. If the authors can provide a clear orientation rule for the eigenvectors, nonnegative edge weights, and an independent safety metric, the paper could become a solid contribution to field robotics."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: TRG-planner is a serious piece of field robotics with a genuinely new graph representation, but Eq. (4) leaves the orientation of the PCA eigenvectors unspecified, and that is a real reproducibility gap in the central risk definition.\n\nWhat I found useful: The Traversal Risk Graph is a reasonable new idea. Modeling each edge's risk via PCA of the local height map, with separate longitudinal and lateral components, captures entry-direction dependence that most traversability maps ignore. The wavefront-based incremental construction and hierarchical management are well designed, and the timing numbers in Table III suggest real-time performance. The ICRA 2023 QRC win is a solid real-world validation, and the qualitative results in Fig. 10, where the robot finds the gentle slope direction, support the direction-aware claim.\n\nThe soft spots are real. First, Eq. (4) defines Rxi_dir as the negative inner product of a PCA eigenvector with the gravity vector. PCA eigenvectors are sign-ambiguous, and the paper never specifies an orientation rule relative to travel direction. Without that, w_ij can flip sign, edge costs in Eq. (7) can go negative, and the A* heuristic becomes inadmissible. The reader's stress-test concern lands. This needs a one-sentence fix, for example orienting e_lon so it points from v_i to v_j or so its z-component is nonnegative, but without it the method is not fully specified as written. Second, gamma and Gamma are tuned on the test scenarios, and the tables report no error bars despite 100 random start-goal pairs, so the performance comparisons are weaker than the text implies. Third, the W metric is computed from the same TRG edge weights the planner minimizes, so it is not an independent safety measure.\n\nNone of this kills the paper. The core idea is sound and the experiments are extensive. The issues are presentation and reproducibility, and they are addressable. I would send this to a serious referee. The right audience is researchers working on traversability-aware planning for legged robots; they will want to cite this for the TRG representation. I would ask the authors to clarify the eigenvector orientation, add variance across the 100 runs, and report at least one independent safety metric.","headline":"Solid field robotics with a genuinely useful graph representation, but Eq. (4) leaves the PCA eigenvector orientation unspecified, which is a real reproducibility gap in the central risk definition.","tokens_in":13443,"tokens_out":1807,"would_cite":true,"duration_ms":18416,"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":"TRG-planner encodes rough terrain as a risk-weighted graph and finds safer, faster paths than A*, PRM*, or T-Hybrid.","keywords":["Path planning","Traversability","Unstructured environments","Legged robots","Field robotics","Traversal Risk Graph","Graph-based planning"],"falsifier":"On the elevation map used in the Long scenario, pick an edge that crosses a slope, evaluate the cost in Eq. (7) twice: once with the principal eigenvector from a standard PCA routine and once with that eigenvector negated. If the resulting route or total cost differs, the risk weight depends on an arbitrary sign convention rather than the terrain geometry.","tokens_in":12439,"feed_emoji":"🤖","tokens_out":9153,"duration_ms":80138,"temperature":0.7,"pith_summary":"The paper proposes TRG-planner, a path-planning method for legged robots in unstructured terrain. Its central claim is that representing the environment as a graph whose edges carry a direction-dependent traversal risk—computed from local slope steepness relative to travel direction—lets a planner find routes that are both safer and shorter than those from A*, PRM*, and T-Hybrid. In 50 m × 50 m simulations with random start-goal pairs, the balanced variant achieves higher travel success and lower path risk than all baselines, with planning times around 3 ms in the longest scenario, roughly an order of magnitude faster than the fastest baseline. The same planner successfully guided a quadruped over a mountain, a one-sided mound, and the ICRA 2023 Quadruped Robot Challenge arena. The method is intended as a global planner that can combine with downstream local optimizers or semantic traversability estimates.","feed_headline":"Risk-weighted graph steers legged robots to safer off-road paths","feed_subtitle":"Edge weights capture slope direction, so a quadruped avoids fall risk that A* and PRM* miss.","key_machinery":"The central object is the Traversal Risk Graph (TRG), a graph G = (V, E) whose nodes are stance patches and whose edges are risk-weighted path candidates. Nodes are sampled outward from the current robot position in a wavefront manner, so only reachable terrain enters the graph; each node stores its 3D position, a validity flag combining geometric stability (height variation below a threshold within a robot-sized inscribed circle) and reachability (at least one incident edge), and its incident edges. Edges are wired between nearby nodes only if the local terrain between them is not too rough or steep; each edge weight is built from a PCA fit of the terrain ellipse around the path, with longitudinal and lateral risk components Rξ_lon and Rξ_lat computed as negative inner products of the principal eigenvectors with gravity. This makes the risk direction-aware: the same slope costs differently when approached from different directions. The graph is extracted locally for updates, expanded from frontier nodes, and integrated back into a global graph; the planning cost in Eq. (7) combines edge distance with the risk weight, enabling A* to trade off distance and safety through the safety factor Γ.","core_discovery":"The central discovery is that relative traversal risk, not just terrain stability, should be the unit of path cost in unstructured environments. TRG-planner builds a Traversal Risk Graph (TRG): nodes are circular terrain patches that the robot can stand on and reach from its current pose, and edges connect nodes whose intervening region is not too steep and whose local height variation is below a threshold. Each edge carries a weight equal to a convex combination of the negative inner products of the two principal PCA direction vectors of the terrain ellipse with gravity, so a path climbing or traversing a slope accumulates risk proportional to its alignment with the fall line. The planner then runs A* on this graph with a cost that sums Euclidean distance plus the risk weight scaled by a safety factor Γ. The paper reports that this representation allows the robot to consistently find paths with lower risk weight, higher travel success, and shorter planning time than conventional methods, and that it worked on a real quadruped in three unstructured field environments.","pith_inferences":["Because edge weights are geometric, they can be combined additively with semantic or learned traversability costs, so the same graph could serve as a global cost skeleton for planners that use image-based terrain classification or learned locomotion costs.","The wavefront construction with frontier-based expansion is naturally suited to streaming elevation maps; a formal incremental variant could update the global graph without rebuilding it, which would reduce replanning latency further as map horizons grow.","The same cost structure could generalize from legged robots to wheeled or tracked platforms by replacing the gravity inner product with a platform-specific rollover or slip risk model, without changing the graph machinery."],"forward_implications":["In the 50 m × 50 m simulation with 100 random start-goal pairs per scenario, the balanced Γ = 3.0 variant beats A*, PRM*, and T-Hybrid in travel success rate for short, medium, and long scenarios, and matches or beats them in path risk weight W.","Planning time is roughly an order of magnitude lower: about 3.00 ms for the long scenario versus 16.89 ms for PRM* and 482.28 ms for T-Hybrid, with no map-preprocessing phase beyond a 4.01 s graph initialization.","The direction-aware risk lets the planner select safe entry directions onto slopes, so the robot climbs a mound only on the side it can actually ascend, and completes the QRC arena course without falling.","The method's behavior can be tuned continuously between distance-minimizing (Γ = 1.0) and risk-minimizing (Γ = 10.0), with the balanced choice producing the best combined travel deviation and success rate."],"supporting_citations":[{"why":"Defines the A* algorithm that serves as the primary baseline and whose cost function the paper extends with traversal risk.","marker":"[34]"},{"why":"Defines PRM*, the asymptotically optimal sampling-based planner used as a comparison baseline.","marker":"[37]"},{"why":"Presents T-Hybrid, the traversability-aware hybrid-map planner compared head-to-head in simulations.","marker":"[24]"},{"why":"Provides the DreamWaQ locomotion controller that lets the quadruped traverse the rough terrain in both simulation and field experiments.","marker":"[6]"},{"why":"Supplies the probabilistic elevation map representation that TRG consumes for local terrain geometry.","marker":"[17]"},{"why":"Introduces the probabilistic roadmap construction that TRG adapts into wavefront-based node sampling and edge wiring.","marker":"[33]"},{"why":"Provides FAST-LIO2 LiDAR-inertial odometry, used to obtain the robot pose and point cloud during real-world experiments.","marker":"[35]"}],"fun_headline_variants":["Risk-weighted graph guides legged robots safer than A* and PRM*","Relative traversal risk sets path cost, not just slope, in TRG-planner","Quadruped wins ICRA 2023 challenge with risk graph path planner","Wavefront-built risk graph yields real-time safe off-road routes","Fall-line risk weights make quadrupeds choose safe, short paths"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The planner's risk value on every edge is built from PCA eigenvectors that are defined only up to a sign, and the paper never fixes their orientation relative to the travel direction, so flipping a sign flips the edge's risk weight and changes the route the cost function produces.","fun_headline_variants_meta":{"raw":{"variants":["Risk-weighted graph guides legged robots safer than A* and PRM*","Relative traversal risk sets path cost, not just slope, in TRG-planner","Quadruped wins ICRA 2023 challenge with risk graph path planner","Wavefront-built risk graph yields real-time safe off-road routes","Fall-line risk weights make quadrupeds choose safe, short paths"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000278,"raw_usage":{"total_tokens":1664,"prompt_tokens":968,"completion_tokens":696,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":599}},"tokens_in":584,"tokens_out":696,"duration_ms":6888,"temperature":1.0,"reasoning_tokens":599,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:20:06.493160+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On the elevation map used in the Long scenario, pick an edge that crosses a slope, evaluate the cost in Eq. (7) twice: once with the principal eigenvector from a standard PCA routine and once with that eigenvector negated. If the resulting route or total cost differs, the risk weight depends on an arbitrary sign convention rather than the terrain geometry.","supporting_citations":[{"cited_title":"A formal basis for the heuristic determination of minimum cost paths,","cited_arxiv_id":null,"evidence_quote":"Defines the A* algorithm that serves as the primary baseline and whose cost function the paper extends with traversal risk."},{"cited_title":"Sampling-based algorithms for optimal motion planning,","cited_arxiv_id":null,"evidence_quote":"Defines PRM*, the asymptotically optimal sampling-based planner used as a comparison baseline."},{"cited_title":"Hybrid map-based path planning for robot navigation in unstructured envi- ronments,","cited_arxiv_id":null,"evidence_quote":"Presents T-Hybrid, the traversability-aware hybrid-map planner compared head-to-head in simulations."},{"cited_title":"Probabilistic terrain mapping for mobile robots with uncertain localization,","cited_arxiv_id":null,"evidence_quote":"Supplies the probabilistic elevation map representation that TRG consumes for local terrain geometry."},{"cited_title":"Prob- abilistic roadmaps for path planning in high-dimensional configuration spaces,","cited_arxiv_id":null,"evidence_quote":"Introduces the probabilistic roadmap construction that TRG adapts into wavefront-based node sampling and edge wiring."},{"cited_title":"FAST-LIO2: Fast direct LiDAR-inertial odometry,","cited_arxiv_id":null,"evidence_quote":"Provides FAST-LIO2 LiDAR-inertial odometry, used to obtain the robot pose and point cloud during real-world experiments."}],"review_version":1}