{"id":"8146cc30-eacc-4024-8baa-3c19d3d0f72a","arxiv_id":"2502.01304","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"An RL policy trained in a new MuJoCo forestry crane simulator grasps varied-diameter logs with 96% simulated success, though the claimed open-source code is not provided.","lead":"The authors built a simulated forestry crane in MuJoCo and trained a reinforcement learning policy to autonomously grasp wood logs of varying sizes. In simulation, the policy succeeds in about 96% of trials, but the claimed open-source code and simulator are not actually released.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (4) defines a target point up to 0.325 m below the log's center, which is below ground for d<0.533 m and biases the reward/observation behind the 96% success claim.","rationale":"The reader's weakest assumption was the ideal velocity controller, which concerns sim-to-real transfer and does not affect the simulation-only claim. The equation issue is more load-bearing because it affects the internal consistency of the method and the trustworthiness of the 96% number even within the simulator. It is also more concrete than the unspecified success thresholds. If Eq. (4) is accurate, the reward/observation are biased below ground for small logs, which should make grasping harder, not easier; the reported high success rate is then suspicious and needs explanation. If it is a typo, it still indicates a quality issue that must be fixed before the benchmark is usable. The verdict remains CONDITIONAL, but the condition should explicitly require correction or justification of Eq. (4) and its use in the reported evaluation.","tokens_in":10429,"tokens_out":8134,"duration_ms":68347,"concrete_test":"Verify Eq. (4) by substituting d=0.3 m (zl=0.15 m) and d=0.8 m (zl=0.4 m) with dmax=0.8 m, and check the target z against the ground plane. Then obtain or reimplement the released code to re-run the Monte Carlo evaluation with the target point corrected to the log's center (Δp = pl - pC). If the success rate drops materially below 96%, the central claim depends on the biased target; if it remains, the authors must explain how the policy compensates for an infeasible observation. If the equation is a typo, correct it and confirm that the published numbers were produced with the corrected version.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Section IV-B.1 defines the log's reduced pose with pl=[xl,yl,zl]^T as the 3D Cartesian position of the log's center point, and Eq. (4) computes the augmented relative distance as Δp = [xl, yl, zl - (dmax - zl)/2]^T - pC(q), with dmax=0.8 m. For a log of diameter d resting on flat ground, zl = d/2, so the target z-coordinate becomes d/2 - (0.8 - d/2)/2 = 0.75d - 0.4. This is negative for all d < 0.533 m, i.e. below the ground plane. Because Δp enters both the observation O = {q, ˙qA, Δp, Δψ} and the reward via dcombine (Eq. 9), the agent is trained against a physically infeasible target for most of the claimed diameter range. The reported 96% success rate therefore cannot be attributed to the stated reward/observation design unless the success threshold for 'misses the center' is extremely lenient or the policy somehow ignores the biased z-component. Either interpretation contradicts the paper's presentation of a geometry-based shaping reward as the driver of the learned skill.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a MuJoCo-based simulation environment for an 8-DOF forestry crane with two unactuated joints, and trains a modified PPO agent (Beta distribution plus uniform perturbation) to grasp cylindrical wood logs of varying diameter (0.3-0.8 m) and random initial poses. In 600 Monte Carlo trials (100 per diameter), the agent achieves a 96.3% success rate in simulation for reaching, grasping, and lifting the log. The paper also benchmarks against TRPO, RecurrentPPO, and PPO, and tests robustness to pose measurement errors.","tokens_in":10720,"tokens_out":6741,"duration_ms":60247,"significance":"If the reported results hold, the simulator and benchmark would provide a useful testbed for RL-based control of large-scale hydraulic manipulators, a domain with few open environments. The Monte Carlo protocol with failure-cause breakdown and the robustness test with pose error are strengths, and the paper is generally clearly written. The novelty of the algorithmic contribution is modest: the Beta distribution policy is known, and the exploration perturbation is adapted from RPO. The central 96% claim is sensitive to a few unspecified evaluation details (Section V-C) and to the construction of the augmented relative distance (Eq. (4)); these issues are addressed in the major comments.","major_comments":[{"comment":"The augmented relative distance defines the target z-coordinate as z_l - (d_max - z_l)/2. For a log of diameter d resting on flat ground, z_l = d/2, so the target z-coordinate becomes 0.75d - 0.4, which is below the ground plane for all d < 0.533 m. Since Delta-p is part of the observation O (IV-B.1) and enters the reward through d_combine in Eq. (9), the policy is trained to drive the grapple center toward a physically infeasible point for most of the claimed diameter range. The manuscript should either justify this offset as an intentional pre-grasp configuration relative to the grapple geometry (with reference to the CAD model) or correct the formula, and it should report the center-miss threshold used in the success criterion (Section V-C) to show that the 96% success rate is not an artifact of this biased target.","section":"Section IV-B.1, Eq. (4)"},{"comment":"The success criterion is not fully specified: the \"particular threshold value\" for missing the center is never given, and \"fully grasp the log\" is not defined in terms of measurable quantities (e.g., joint angle q8 threshold, contact forces, or lift height). Without these definitions, the reported success rates in Tables I and II are not reproducible. This is load-bearing because the policy is trained with the same d_combine metric used in the early-termination criterion (IV-B.3), so the evaluation could partially measure optimization of the training objective rather than actual grasp quality.","section":"Section V-C"},{"comment":"The paper states that it provides \"an open-source benchmark for the community,\" but no code, simulator, or dataset repository is provided; the only external link is a demonstration video. For a paper whose central contribution is a simulator and benchmark, the absence of a code release link makes the claim unverifiable and the contribution incomplete.","section":"Abstract and Section I"}],"minor_comments":[{"comment":"Typo: \"sim-to-read gap\" should be \"sim-to-real gap\".","section":"Section V-D"},{"comment":"The reward weights omega1, omega2, and omega3 in Eqs. (9)-(12) are user-defined parameters but their values are never reported; these should be given for reproducibility.","section":"Section IV-B.2"},{"comment":"The early-termination threshold epsilon in the criterion \"d_combine < epsilon\" is not specified.","section":"Section IV-B.3"},{"comment":"The description of the randomized log region is vague (\"the center of this region is approximately 6.5 m from the crane's base\"); please provide the region's geometry or bounds.","section":"Section III-B"},{"comment":"The spelling of \"MuJoCo\" is inconsistent (e.g., \"Mujoco\" in the abstract and Section V).","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The Eq. (4) issue and the unspecified success threshold are the key concerns; they should be resolved before the benchmark claims can be accepted. The missing code repository is also a notable gap for a benchmark paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: this paper is worth a look if you care about heavy-duty manipulation, but the headline number is softer than it looks.\n\nThe new thing is the simulator: a CAD-based MuJoCo model of an 8-DoF forestry crane with variable log diameters (0.3–0.8 m), plus a benchmark of PPO variants against TRPO and recurrent PPO. That fills a real gap; prior RL work used fixed-diameter logs and commercial dynamics software. The mPPO tweak (Beta distribution plus a uniform perturbation) is modest, but the comparison is honest and the Monte Carlo evaluation is substantial: 100 trials per diameter across six diameters, plus a robustness test with injected pose errors. That part is solid.\n\nThe soft spots are real. Equation (4) defines the target for the grapple center as [xl, yl, zl - (dmax - zl)/2]. For a log resting on flat ground, zl = d/2, so this target z is 0.75d - 0.4, which is below grade for any diameter under 0.533 m. For the smallest logs, the target sits 0.175 m underground. The paper presents this offset as a sensible size-dependent shaping term, but it is physically infeasible. Either the success threshold for \"missing the center\" is loose enough to tolerate the bias, or the policy learns to ignore the z-component; neither reading matches the text. And that threshold is never specified. This is the main thing to fix before trusting the 96% number.\n\nSecond, the paper advertises an open-source benchmark, but I find no released code or simulator. A benchmark paper that doesn't ship the benchmark is half a benchmark. Third, the ideal velocity controller assumption is stated as a limitation, but its effect on sim-to-real transfer is not quantified. That is acceptable for a simulation study, but it should temper the \"towards autonomous\" framing.\n\nThis is not a reject-able paper. The simulator and task definition are genuinely useful, and the authors are transparent about limitations. But the z-offset issue is load-bearing for the reward design, and the missing threshold makes the headline claim hard to interpret. I would send it out for review because the environment is a real contribution and the flaws are addressable, but I would expect major revisions.","headline":"Useful simulator and benchmark for forestry-crane grasping, but the 96% success claim rests on an unspecified threshold and a reward offset that points below ground for most log sizes.","tokens_in":11252,"tokens_out":5908,"would_cite":false,"duration_ms":50571,"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":"A reinforcement-learning controller trained in a new MuJoCo simulator grasps and lifts varied-diameter wood logs with a forestry crane at a success rate above 96% in Monte Carlo trials.","keywords":["forestry crane","reinforcement learning","log grasping","MuJoCo simulator","proximal policy optimization","latent Markov decision process","benchmark","sim-to-real"],"falsifier":"Run the trained mPPO policy on the physical forestry crane using the same Monte Carlo protocol (six log diameters, 100 trials per diameter, randomized initial configurations) and compare the success rate to the reported 96%. A more direct measurement is to record the actual joint velocity tracking error of the hydraulic actuators during a grasping sequence and check whether it exceeds the bounds implicitly assumed by the ideal velocity controller in the simulator.","tokens_in":10263,"feed_emoji":"🪵","tokens_out":9001,"duration_ms":74944,"temperature":0.7,"pith_summary":"The paper makes the case that an autonomous forestry crane can learn to grasp and lift heavy wood logs of varying diameter entirely from reinforcement learning in a custom MuJoCo simulator. It reports that the trained mPPO policy succeeds in over 96% of Monte Carlo trials spanning log diameters from 0.3 to 0.8 m and randomized initial crane configurations, and still about 92% of trials when the log-pose input is corrupted with measurement error. The contribution is a reproducible open-source benchmark: the simulator built from CAD data, the reward functions, and the modified PPO algorithm. The result matters because forestry cranes are hydraulically actuated, underactuated, and currently operated manually by a shrinking pool of skilled operators; a simulator-trained controller that works would make automation far cheaper and safer to develop.","feed_headline":"Sim-trained forestry crane grasps 96% of varied logs","feed_subtitle":"Open MuJoCo simulator and a modified PPO let an 8-DoF crane lift logs of any diameter from 0.3 to 0.8 m.","key_machinery":"The load-bearing machinery is the MuJoCo simulator of the 8-DoF forestry crane built from CAD data, with 38 rigid bodies, two pairs of synchronized joints, and signed-distance-field collision primitives for the grapple jaws; the latent-MDP training setup where log diameter is a latent variable sampled uniformly across episodes; and the mPPO algorithm, a variant of PPO that uses a Beta distribution for the policy (so actions inherently respect joint-velocity limits) and perturbs sampled actions with uniform noise $g \\sim U(-\\epsilon,\\epsilon)$ clipped to $[0,1]$ (with $\\epsilon=0.1$) to aid exploration. The reward function is the fourth component: it sums distance, grapple, lift, and balance terms, each designed to dominate at a different stage of the grasping task.","core_discovery":"On its own terms, the central discovery is that a model-free RL policy can perform the complete wood-log grasping cycle—approaching, aligning the grapple, closing on the log, lifting it, and stabilizing the load—in simulation with a success rate exceeding 96% across logs of diameter 0.3–0.8 m and random crane initializations. The key modification is mPPO, which replaces the Gaussian policy distribution of standard PPO with a Beta distribution so that commanded joint velocities always lie within the admissible range, and which adds a small uniform perturbation to sampled actions during training to encourage exploration in the large search space. The policy is trained under the latent-MDP formulation in which the log diameter is a latent variable sampled uniformly, forcing the agent to serve all diameters without retraining. The reward function decomposes the task into four sequenced terms (distance, grapple, lift, balance), and the paper shows that the same reward and simulator produce stronger cumulative reward with mPPO than with standard PPO, TRPO, or Recurrent PPO. When the assumed log pose is perturbed by up to ±10% measurement error, the success rate drops only to about 92%, indicating some tolerance for perception noise.","pith_inferences":["The 96% figure is a simulation number obtained under ideal velocity control; a real deployment will likely require domain randomization on hydraulic parameters to keep the success rate, and the paper's own limitation section acknowledges this as the main open gap.","Because the policy is trained with the log pose given, the practical bottleneck in the field is probably the pose-estimation system; the ±10% noise test implies the perception pipeline only needs to keep relative pose errors within that range for success.","The latent-MDP framing with diameter as the latent variable should extend to other varying log properties, such as length or mass, as long as they are sampled during training, which could remove the fixed-length restriction in the current simulator.","A direct comparison between the ideal velocity controller and a simulated hydraulic actuator model (with pump flow limits and delays) in the same MuJoCo scene would quantify how much of the promised performance depends on that assumption."],"forward_implications":["A single mPPO policy, trained once, covers the full log-diameter range from 0.3 to 0.8 m without retraining, because the latent-MDP formulation treats diameter as a latent variable.","The benchmark defines a fixed success protocol (reach within 6 s, grasp fully within 9 s, grasp near the log's center) that other researchers can adopt for quantitative comparison.","Adding up to ±10% relative measurement error to the log pose lowers success from >96% to approximately 92%, giving a concrete accuracy target for the perception system.","The mPPO variant runs at roughly 6500 simulation steps per second, about four times faster than TRPO and Recurrent PPO, making large-scale training practical on a single desktop GPU."],"supporting_citations":[{"why":"Closest prior RL baseline for forestry crane log grasping with a fixed-size log; provides the task and baseline this work extends to varied diameters.","marker":"[14]"},{"why":"The PPO algorithm that mPPO modifies and that serves as the primary baseline in the benchmark.","marker":"[19]"},{"why":"The MuJoCo physics engine that supplies the simulator's dynamics and multi-contact simulation.","marker":"[23]"},{"why":"Along with [26], provides the latent-MDP formulation used to treat log diameter as a latent variable.","marker":"[25]"},{"why":"The Beta-distribution policy for continuous control that underlies mPPO's action distribution.","marker":"[28]"},{"why":"The open-source PPO implementation that mPPO builds on, enabling the reproducible benchmark.","marker":"[33]"}],"fun_headline_variants":["RL crane lifts 96% of logs in MuJoCo sim","Forestry crane RL hits 96% grasp success","Autonomous log grasping: 96% in simulator","Beta-policy crane grasps 96% of varied logs","Sim crane scores 96% on log grasp benchmark"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes an ideal underlying velocity controller for the six actuated joints, neglecting hydraulic actuator dynamics, delays, and pump flow limits; if the real crane's joints cannot track the commanded velocities closely, the simulator-trained policy's success rate could drop substantially.","fun_headline_variants_meta":{"raw":{"variants":["RL crane lifts 96% of logs in MuJoCo sim","Forestry crane RL hits 96% grasp success","Autonomous log grasping: 96% in simulator","Beta-policy crane grasps 96% of varied logs","Sim crane scores 96% on log grasp benchmark"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000272,"raw_usage":{"total_tokens":1647,"prompt_tokens":976,"completion_tokens":671,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":592,"completion_tokens_details":{"reasoning_tokens":590}},"tokens_in":592,"tokens_out":671,"duration_ms":6348,"temperature":1.0,"reasoning_tokens":590,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T15:44:44.934075+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained mPPO policy on the physical forestry crane using the same Monte Carlo protocol (six log diameters, 100 trials per diameter, randomized initial configurations) and compare the success rate to the reported 96%. A more direct measurement is to record the actual joint velocity tracking error of the hydraulic actuators during a grasping sequence and check whether it exceeds the bounds implicitly assumed by the ideal velocity controller in the simulator.","supporting_citations":[{"cited_title":"Reinforcement learning control of a forestry crane manipulator,","cited_arxiv_id":null,"evidence_quote":"Closest prior RL baseline for forestry crane log grasping with a fixed-size log; provides the task and baseline this work extends to varied diameters."},{"cited_title":"Mujoco: A physics engine for model-based control,","cited_arxiv_id":null,"evidence_quote":"The MuJoCo physics engine that supplies the simulator's dynamics and multi-contact simulation."},{"cited_title":"Understanding domain randomization for sim-to-real transfer,","cited_arxiv_id":null,"evidence_quote":"Along with [26], provides the latent-MDP formulation used to treat log diameter as a latent variable."},{"cited_title":"Improving stochastic policy gradients in continuous control with deep reinforcement learning using the beta distribution,","cited_arxiv_id":null,"evidence_quote":"The Beta-distribution policy for continuous control that underlies mPPO's action distribution."},{"cited_title":"Stable-baselines3: Reliable reinforcement learning im- plementations,","cited_arxiv_id":null,"evidence_quote":"The open-source PPO implementation that mPPO builds on, enabling the reproducible benchmark."}],"review_version":1}