{"id":"d4e82c30-a43e-487d-b23c-b6af7a1904ee","arxiv_id":"2501.14942","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"Force-based human demonstrations, collected through a haptic VR-robot setup, improved simulated pipe-insertion policy learning over visual demonstrations.","lead":"This paper builds a two-phase system that records a person's force-guided pipe insertion motions in a simulated environment, then uses those demonstrations to train a simulated robot with imitation plus reinforcement learning. The authors report that force-based demonstrations trained the robot faster and with higher success rates than visual demonstrations in the simulated task.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The force group's edge is confounded: Eq. (11) rewards decreasing collision forces that only the force observation provides, so the comparison may show reward-observation alignment rather than a benefit of force demonstrations.","rationale":"The reader's formal weakest_assumption is the realism of the custom Unity force simulation and lack of real-world validation. That is an important external-validity concern. However, the more load-bearing problem for the paper's central claim is internal to the reported experiments: the force group is given observations that exactly match the terms rewarded in Eq. (11), while the visual group is not. This makes the force group's optimization problem easier regardless of whether force is intrinsically a better demonstration modality. The reader's rationale does mention this Eq. (11) confound, but their designated weakest assumption was the simulation realism, so my agreement is partial. I recommend keeping the reader's REJECT verdict unchanged: the reward confound alone invalidates the headline comparison as stated, and the unvalidated simulator adds a second, separate reason the claim is not established.","tokens_in":17974,"tokens_out":5023,"duration_ms":48427,"concrete_test":"Retrain both groups with the same demonstrations and observation spaces, but replace Eq. (11) with a task-progress reward that does not use force magnitudes (e.g., success bonus plus negative per-step penalty based only on distance/depth). If the force group's advantage in success rate and episode length disappears or shrinks materially, the claimed benefit of force demonstration is an artifact of reward-observation alignment rather than a property of force feedback.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central comparison in Sections 4 and 5 varies three things at once: demonstration modality, agent observation space, and the reward's dependence on the force observation. In Section 3.4.2 (deep learning process), the reward r_t in Eq. (11) gives positive credit exactly when ||F_normal^t|| and ||F_friction^t|| decrease relative to the previous step. These are precisely the two components of the force observation vector F_obs. The visual condition's observation is Vis = {d, l, [P]} from 64 Raycasters and contains no force channel, so the visual agent is rewarded for reducing quantities it cannot perceive. PPO can partially infer them from images, but the force group receives the optimization target directly in its state, making its learning problem substantially easier. The paper's conclusion that force demonstrations outperform visual demonstrations therefore conflates the demonstration channel with a reward-observation alignment advantage. The unvalidated custom Unity contact model (Section 3.4) is a separate external-validity concern; even within the simulation, this reward confound is sufficient to undermine the headline result as stated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-phase system for learning pipe-insertion assembly: in the first phase, a human operator guides a Franka Emika arm that is synchronized with a Unity simulator, receiving haptic feedback from a custom collision/force model; in the second phase, the recorded demonstrations are used to initialize a policy via GAIL, which is then fine-tuned with PPO. The authors compare agents trained on force observations (F_obs) against agents trained on ray-caster visual observations (Vis_t) and against reward-only RL baselines, reporting faster convergence, higher success rates, and shorter episode lengths for the force-based agents across 20 demonstrations per condition. The abstract and conclusion claim that force feedback is a superior demonstration modality for contact-rich construction assembly tasks.","tokens_in":18209,"tokens_out":4605,"duration_ms":43990,"significance":"If the central comparison were valid, the paper would provide a useful practical contribution: an end-to-end human-in-the-loop data-collection and learning pipeline that uses force feedback for contact-rich assembly, with a moderately large set of demonstrations and 100 inference trials per learned policy. The authors also explicitly examine within-group variability in Section 5.4, which is a commendable check on the robustness of their results. However, the headline claim is currently undermined by a methodological confound: the reward in Eq. (11) directly rewards reductions in the very quantities that appear only in the force observation, so the reported advantage may reflect reward-observation alignment rather than any benefit of force demonstrations as a modality. The custom Unity force model is also unvalidated against physical measurements, which limits the external validity of any simulation-only conclusion. With the confounds removed or controlled and with statistical inference added, the question would be well worth revisiting.","major_comments":[{"comment":"The central comparison is confounded. The reward r_t in Eq. (11) gives positive feedback exactly when ||F_normal^t|| and ||F_friction^t|| decrease relative to the previous step, and the force observation F_obs is defined as [F_normal, F_friction]. The visual observation Vis_t = {d, l, [P]} contains no force channel, and the RL baseline observes neither forces nor ray-caster points. Consequently, the force-condition agent is rewarded for reducing quantities that are directly in its state, while the visual and RL conditions are rewarded for reducing quantities they cannot perceive. This means the experiments vary at least two things at once: demonstration modality (force vs. visual) and alignment between the reward and the agent's observation. The paper's conclusion in Sections 5 and 6 that force demonstrations outperform visual demonstrations therefore does not follow from the reported results. I would need an ablation that fixes the reward-observation alignment, for example by giving the visual condition the same force-decrease reward computed from the simulator's hidden state, or by removing the force-decrease component from the force condition's reward, or by training a force-observation agent on visual/hand-crafted demonstrations.","section":"Section 3.4.2, Eq. (11)"},{"comment":"The external validity of the force model is unsupported. The simulator's contact forces are produced by a custom impedance-control and impulse model whose parameters (mass m, stiffness K_d, ray count n_rays, 45-degree threshold) are chosen without calibration against any physical force-torque measurement. The paper states in Section 6 that validation with real-world data is future work, but the abstract and introduction describe the simulation as providing 'realistic' or 'similar-to-real' force interactions. Since the demonstrations, the agent observations, the reward, and the success criterion all come from this same unvalidated model, a reader cannot determine whether the learned force-based policies would transfer to a physical pipe-insertion task. I recommend either calibrating the model against measured contact forces or substantially tempering the real-world claims and explicitly labeling the results as simulation-only.","section":"Section 3.4 and Section 6"},{"comment":"The superiority claim is reported without any statistical inference. Each trial yields a success rate and mean episode length over 100 inferences, but the paper only reports means and standard deviations across the 20 demonstrations and does not test whether the force-group distribution differs from the visual-group distribution. This matters because Section 5.4 identifies two force demonstrations (8 and 20) whose inference success rates are below their visual counterparts; with 20 paired samples and this level of variability, the headline difference could be driven by a few large successes. I would like to see paired per-demonstration differences with a nonparametric test (e.g., Wilcoxon signed-rank), confidence intervals for success rates, and an explicit statement about how many of the 20 force demonstrations beat their visual counterpart.","section":"Section 5.2, 5.3, and 5.4, Tables 2-5"}],"minor_comments":[{"comment":"The visual-group table for randomized inner pipe location is labeled 'Table 4' but should be Table 3, and the cross-references in Sections 5.2 and 5.3 are inconsistent; please renumber all tables and fix the in-text references.","section":"Section 5.2, Tables 3 and 4"},{"comment":"In Eq. (1), K_d is described as a 'damping coefficient for motion speed,' but it appears as the coefficient of the displacement term in an impedance control law and should be labeled as a stiffness coefficient; please clarify the physical meaning and units of M_d, B_d, and K_d.","section":"Section 3.3, Eq. (1)"},{"comment":"The sign convention for the indicator Ind in Eq. (10) appears inconsistent with the earlier definition of delta_move in Eq. (2): if Ind >= 0, the paper sets delta_c to 0, yet a positive dot product would normally mean the handler is still pushing toward the collision point; please reconcile the sign convention and the verbal description.","section":"Section 3.4.2, Eq. (10)"},{"comment":"The paper uses promotional language such as 'groundbreaking' and contains typographical errors such as 'EXPEERIMENTS'; I recommend a careful language edit and neutral terminology throughout.","section":"Section 1 and Section 4"}],"recommendation":"major_revision","confidential_remarks":"I agree with the reader's assessment that the reward-observation confound is real and load-bearing; however, I do not see it as an unfixable error. The authors could address it with targeted ablation experiments and paired statistical tests, and those additions lie within the normal scope of a revision. The unvalidated force model is a separate external-validity concern that should at minimum be reflected in the paper's claims. I would not accept the manuscript in its current form, but I would not reject it outright if the authors are willing to perform the necessary controls."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, the short version: this is a genuine systems-building effort for force-based imitation learning in construction assembly, but the comparison that carries the paper is confounded, and the headline result is not established.\n\nWhat is actually new: the authors built a two-phase pipeline where a human moves a Franka arm with a pipe handle, sees a Unity twin through VR, feels collision forces via haptic feedback computed by a custom trigger-based contact model, and the resulting demonstrations train GAIL+PPO policies. That integration—haptic VR data collection loop plus force-conditioned GAIL—is not something I have seen in the construction robotics literature. They also did more than a toy demo: 20 demonstrations per modality, randomized pipe and target positions, and 100 inference trials per policy. The citation pattern is fine; they locate themselves in the construction IL line and cite GAIL correctly.\n\nThe soft spots are serious. The central claim is that force demonstrations beat visual demonstrations. But Eq. (11) rewards the agent whenever normal and friction forces decrease, and those quantities are exactly the force observation. The visual condition does not observe forces, so the visual agent is rewarded for reducing information it cannot perceive. The comparison therefore varies demonstration modality, observation space, and reward signal at once. That is a load-bearing confound. The force group's advantage may simply be reward-observation alignment, not a property of force demonstrations per se. The stress-test note is correct.\n\nSecond, all training and evaluation happen inside the authors' custom Unity force model. The contact forces come from impedance control with a large stiffness, a predefined mass, ray-cast SVD normals, and trigger-based detection, but there are no real force-torque measurements to validate against. The authors acknowledge this and defer real-world validation to future work, but without it the \"similar-to-real\" claim is unsupported and transfer remains speculative. No code or data are released, and there are no multiple training seeds for the same demonstration, so the variance numbers are thinner than they look.\n\nWhat the paper does well: the framework is coherent, the writing is readable, and the authors are honest about the variability in two of their force demonstrations, which is more than many papers do. But the empirical core does not support the stated conclusion.\n\nThis paper is for researchers working on haptic demonstration collection or force-conditioned imitation learning for contact-rich construction tasks. It would be a useful systems contribution if the comparison were fixed. I would not cite it as evidence that force demonstrations outperform visual ones. My recommendation: do not publish as is. A serious venue could send it to review only with the expectation of a reworked, modality-fair baseline, real force validation, and released artifacts; otherwise it should be rejected.","headline":"A real haptic demonstration pipeline for construction assembly, but the force-vs-visual comparison is confounded by a reward only the force group can perceive.","tokens_in":18730,"tokens_out":3661,"would_cite":false,"duration_ms":50581,"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":"Force feedback from human demonstrations, distilled through GAIL and PPO, beats vision and pure RL for simulated pipe insertion.","keywords":["imitation learning","force feedback","haptic demonstration","construction robotics","pipe insertion","generative adversarial imitation learning","proximal policy optimization","digital twin"],"falsifier":"Replay the recorded human motions on a physical pipe-insertion rig instrumented with a force-torque sensor and compare the measured normal and friction profiles with what the Unity simulation outputs for the same trajectories; then train the GAIL-PPO pipeline on real-force demonstrations and run the learned policy on the physical robot. If the simulated contact forces differ substantially from measured ones, or if policies trained on real forces do not approach the reported 74-81% success rates, then the force-modality advantage is a property of the simulator, not of force feedback as such.","tokens_in":17733,"feed_emoji":"🤖","tokens_out":17383,"duration_ms":124631,"temperature":0.7,"pith_summary":"The paper sets out to show that force feedback, not vision, should carry imitation learning for contact-rich construction tasks such as pipe insertion, where the insertion point is occluded and surface irregularities defeat visual cues. Its two-phase framework first records human demonstrations on a robot arm whose Unity digital twin generates haptic collision feedback, so the demonstrations themselves contain human force responses, and then converts those demonstrations into robot policy via Generative Adversarial Imitation Learning (GAIL) initialization refined by Proximal Policy Optimization (PPO), with normal and friction forces folded into the reward. The reported result is that force-trained policies complete simulated pipe insertion faster and more reliably: average success of 81.2% and 74.1% under two randomization conditions, against 27.6% and 25.7% for visual demonstration training and 24% and 14% for pure reinforcement learning. If the result holds beyond simulation, construction robots could learn dexterous assembly from human touch rather than from cameras.","feed_headline":"Force-based training beats vision for simulated pipe insertion","feed_subtitle":"In simulation, pipe-insertion policies from human force demos succeed at 74–81%, versus 26–28% for visual demos.","key_machinery":"The mechanism that carries the argument is the two-phase DCI-LG loop with its custom contact-force core. In the Data Collection through Interaction (DCI) phase, a human pushes a pipe-handle mounted on a 6-DoF Panda arm while ROS-Sharp streams the pose into a Unity digital twin; contact is detected with trigger volumes instead of the full physics solver, and the impedance controller $F = \\delta_{coll}\\delta_{dir}(M_1\\ddot{x}_{dist} + B_1\\dot{x}_{dist} + K_1 x_{dist}) + g_c$ with a very large stiffness $K_1$ turns virtual collisions into haptic resistance on the real arm. The direction filter $\\delta_{dir}$ is the piece that makes the force feel physical: resistance is applied only while the operator is pushing toward the last collision point, and it vanishes when the operator pulls away. Collision impulses are scaled by a predefined pipe mass $m$ and frame-to-frame velocity, the contact normal comes from a raycast fan around the closest contact point whose covariance is decomposed by singular value decomposition, and the resulting normal force $\\mathbf{F}_{normal}$ and friction $\\mathbf{F}_{friction}$ form the force observation $\\mathbf{F}_{obs}$. In the Learning with Guidance (LG) phase, this observation plus depth and distance enters a GAIL generator with a 256-neuron hidden layer that emits a 3-D Cartesian force action, and the policy is then fine-tuned with PPO; the reward $r_t$ pays positive score when both normal and friction forces shrink across steps and adds +1 when the pipe reaches the 0.5 m target depth, which is what presses the agent toward insertion through decreasing contact rather than brute force.","core_discovery":"The paper's central claim is that force feedback is a better demonstration modality than vision for training contact-rich construction assembly, and that the quality of force demonstrations depends on letting the human operator feel realistic contact while collecting them. The authors implement this claim as a two-phase system: humans first guide a physical robot arm through pipe insertion while a Unity-simulated twin reflects collisions back as haptic resistance, so the 20 recorded demonstrations per condition encode real human correction behavior under contact; then Generative Adversarial Imitation Learning (GAIL) converts those demonstrations into an initial policy, which Proximal Policy Optimization (PPO) refines, with normal and friction forces embedded in the observation and reward. On inference tests with randomized pipe positions, force-trained agents succeed at an average of 81.2% over 100 trials per demonstration versus 27.6% for visually trained agents, and the gap persists (74.1% versus 25.7%) when the target location is also randomized; pure reinforcement learning without force observations reaches only 24% and 14% in the two conditions. The authors further report that force demonstrations converge faster in training and produce narrower reward distributions, though they acknowledge two outlier demonstrations that underperform their visual counterparts.","pith_inferences":["Because the paper never measures real contact forces, the decisive extension it implies but does not run is a sim-to-real check: replay recorded human motions on an instrumented physical insertion rig and compare measured normal and friction profiles with the simulated ones; the reward function's assumptions stand or fall on that match.","The reward design, paying positive score for shrinking normal and friction forces, is stated only for pipe insertion, but the same shaping should apply to any peg-in-hole or alignment-through-contact task, so the claimed benefit is plausibly a general recipe for contact-rich assembly rather than a pipe-specific trick; the paper does not test that generality.","The two outlier demonstrations suggest a practical selection rule the paper only gestures at: filter or reweight demonstrations by force-profile consistency before GAIL initialization, which could remove the policies that drag the average down.","A data-efficiency argument is implicit in the architecture: the force observation is eight numbers versus 258 for the visual encoder, so if the success-rate gap holds, force feedback is also the cheaper demonstration modality; quantifying the sample-complexity difference would make that case explicit."],"forward_implications":["Force-guided trials converge faster than visual-guided trials for all 20 demonstrations per group, so training time is cut when force demonstrations are used.","Force-trained policies keep most of their edge under randomization: 74.1% average success when both pipe positions are randomized, versus 25.7% for visual policies, indicating the force observation carries transferable alignment information.","Imitation initialization adds to the force feedback: the pure-RL condition (position-only observation) converges to roughly 70% of the force-feedback reward, and every force demonstration trained with the DIL-DRL workflow outperforms the same force observations trained with DRL alone, so both the force modality and the imitation step contribute.","Force demonstrations produce narrower performance distributions, meaning more consistent policies, though two of the twenty force demonstrations (trials 8 and 20) converge slower and generalize worse than their visual counterparts."],"supporting_citations":[{"why":"Supplies the Generative Adversarial Imitation Learning algorithm whose discriminator reward is the core of the learning-with-guidance phase.","marker":"[15]"},{"why":"Shows that demonstration data generated in a virtual environment can train construction robots, the premise the data-collection rig builds on and extends to force.","marker":"[46]"},{"why":"The visual-GAIL construction framework that serves as the comparison baseline the force approach is claimed to beat.","marker":"[73]"},{"why":"Demonstrates tactile and force feedback as a usable interaction signal in construction, motivating the choice of force demonstrations.","marker":"[79]"},{"why":"Justifies the DIL-initialized-then-DRL-refined training workflow used in the learning phase.","marker":"[52]"},{"why":"Arguing that imitation alone is insufficient motivates the combined DIL-DRL training that the paper adopts.","marker":"[47]"},{"why":"The bi-directional digital-twin synchronization concept that underpins the ROS-Sharp data bridge between the real arm and Unity.","marker":"[56]"}],"fun_headline_variants":["Force feedback demos triple success in simulated pipe insertion","Haptic training outperforms visual demos 3-to-1 for robot assembly","Robots learn better by touch than sight for construction tasks","Human force demos teach robots pipe insertion 3x better than vision","Force-based imitation learning beats visual for pipe insertion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole framework rests on the assumption that the Unity simulation's contact forces, including stiffness resistance, collision impulses, and computed surface normals, feel and behave enough like real pipe-insertion forces that human adjustments collected in the rig and the policies trained on them transfer to a physical robot, and the paper reports no real force measurements to validate that.","fun_headline_variants_meta":{"raw":{"variants":["Force feedback demos triple success in simulated pipe insertion","Haptic training outperforms visual demos 3-to-1 for robot assembly","Robots learn better by touch than sight for construction tasks","Human force demos teach robots pipe insertion 3x better than vision","Force-based imitation learning beats visual for pipe insertion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000316,"raw_usage":{"total_tokens":1778,"prompt_tokens":922,"completion_tokens":856,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":538,"completion_tokens_details":{"reasoning_tokens":770}},"tokens_in":538,"tokens_out":856,"duration_ms":8152,"temperature":1.0,"reasoning_tokens":770,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T14:46:27.241089+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replay the recorded human motions on a physical pipe-insertion rig instrumented with a force-torque sensor and compare the measured normal and friction profiles with what the Unity simulation outputs for the same trajectories; then train the GAIL-PPO pipeline on real-force demonstrations and run the learned policy on the physical robot. If the simulated contact forces differ substantially from measured ones, or if policies trained on real forces do not approach the reported 74-81% success rates, then the force-modality advantage is a property of the simulator, not of force feedback as such.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Generative Adversarial Imitation Learning algorithm whose discriminator reward is the core of the learning-with-guidance phase."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows that demonstration data generated in a virtual environment can train construction robots, the premise the data-collection rig builds on and extends to force."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The visual-GAIL construction framework that serves as the comparison baseline the force approach is claimed to beat."},{"cited_title":"Yu, V .R","cited_arxiv_id":null,"evidence_quote":"Demonstrates tactile and force feedback as a usable interaction signal in construction, motivating the choice of force demonstrations."},{"cited_title":"Sasagawa, K","cited_arxiv_id":null,"evidence_quote":"Arguing that imitation alone is insufficient motivates the combined DIL-DRL training that the paper adopts."},{"cited_title":"Wang, C.-J","cited_arxiv_id":null,"evidence_quote":"The bi-directional digital-twin synchronization concept that underpins the ROS-Sharp data bridge between the real arm and Unity."}],"review_version":1}