{"id":"bb651ac7-6cb1-4416-a3c5-a4defa442932","arxiv_id":"2607.11734","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A Transformer-based actuator model learns torque surrogates and external forces on low-cost servo arms from pose trajectories and motor telemetry, improving force estimation and behavior-cloning control.","lead":"A team at MIT trains a Transformer to replace the unreliable linear current-to-torque model in low-cost robot arms, and to estimate grip forces without a force sensor. The system includes a new teleoperation dataset across three robot platforms and shows better force estimates and imitation-learning success.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Implicit coupling leaves the force head as a telemetry-to-force regressor: f_hat never enters the dynamics, so physically grounded sensorless force perception is not established outside the labeled task set.","rationale":"The reader's verdict is already CONDITIONAL and identifies the same load-bearing assumption. My stress-test confirms it: implicit coupling is the paper's own chosen mode (Sec. IV-L), and the paper itself flags the identifiability issue in Sec. III-B. The force head is trained purely by regression to labels, so the 'jointly predicts torque and force' formulation does not by itself make the force estimator physical. The real-robot dataset, public release, and multi-platform evaluation are genuine strengths, and the paper is unusually transparent about the distinction between recorded-telemetry-conditioned rollouts and counterfactual forecasting. Still, because the central contribution includes sensorless force perception, a direct learned-current-to-force baseline and out-of-distribution force evaluation are needed before the physical-grounding claim can be accepted. This does not change the conditional verdict; it specifies the condition.","tokens_in":32434,"tokens_out":6228,"duration_ms":68292,"concrete_test":"Train a direct learned current-to-force regressor (e.g., an MLP over the same NAD telemetry features: current/effort, q, q_cmd, and history) with the same payload labels, but no differentiable simulation or torque surrogate. Compare force MAE on the Tab. V payload benchmark and on held-out payloads (e.g., 150 g and 600 g dynamic lifts) and non-gravitational pushes. If the regressor matches or beats NeuralActuator, the simulator/torque-surrogate branch is not load-bearing for force perception and the physical-grounding claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In the deployed implicit-coupling mode (Sec. III-B, Eq. 11), tau_pred alone advances the simulated state while f_hat_ext is supervised but never applied to the simulator. The paper explicitly concedes: 'The torque surrogate may therefore absorb interaction effects needed to reproduce the observed motion.' Consequently the force head is not constrained by the rigid-body residual tau_ID - tau_act of Eq. 4; it is a supervised telemetry-to-force regressor whose predictions do not affect the dynamics on which the torque surrogate is trained. The strong force MAEs in Tabs. IV-VI may therefore reflect learned correlations between motor current/task identity and the constant gravity labels of the payload protocol, rather than a dynamics-based estimate. The two 'unseen contact geometry' checks are stationary holds, and the Franka benchmark is explicitly future-state-conditioned (App. B), so neither establishes causal generalization outside the labeled task set. This is load-bearing because the headline contribution is 'force perception without force/torque sensors at inference time,' a claim that requires the force head to be more than a task-set interpolator.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces NeuralActuator, a Transformer-based multi-task model that predicts (i) a torque surrogate used to advance a differentiable rigid-body simulator, (ii) a contact-gated external force estimate, and (iii) per-motor condition scores. The torque-surrogate head is trained from pose trajectories through differentiable simulation, avoiding direct joint-torque labels. The authors collect a synchronized teleoperation dataset (NAD) on OpenManipulator-X, extend to SO-101 and Franka, and report low rollout errors, force MAEs of 0.08–0.12 N on payload benchmarks and 0.47–0.73 N on SO-101, improved behavior-cloning success rates, and sub-millisecond inference. The paper is transparent about several limitations, including the implicit force-coupling setup, recorded-telemetry-conditioned rollouts, and the future-state-conditioned Franka benchmark.","tokens_in":32782,"tokens_out":4083,"duration_ms":43881,"significance":"If the force-estimation results hold under causal deployment and consistent evaluation, NeuralActuator would be a valuable contribution to low-cost robot actuation modeling and sensorless force perception. The released dataset, code, and hardware configurations, together with cross-platform validation, careful ablations, and real-time performance analysis, are notable strengths. The central ‘sensorless force perception’ claim, however, currently rests on implicitly coupled, telemetry-conditioned evaluations with a training/evaluation label-alignment mismatch and nominal payload labels during motion. Additional evidence is needed to establish that the force head is physically grounded and generalizes beyond the labeled task set. The work is not circular—the force head is supervised by independent force labels—but the identifiability and generalization concerns are load-bearing for the headline claim.","major_comments":[{"comment":"The deployed implicit coupling lets the torque surrogate alone drive the simulator while the force head is supervised but never applied to the dynamics. As the paper states, “The torque surrogate may therefore absorb interaction effects needed to reproduce the observed motion.” Consequently, the force head is not constrained by the residual τ_ID − τ_act of Eq. (4); it is a telemetry-to-force regressor whose strong MAEs in Tabs. IV–VI may reflect correlations between motor current/task identity and the constant or near-constant labels of the protocol, rather than a dynamics-based estimate. The two “unseen contact geometry” checks are stationary holds with constant loads, and the Franka benchmark is explicitly future-state-conditioned, so neither establishes causal generalization outside the labeled task set. Please provide a causal evaluation of the force head under explicit coupling, or","section":"Sec. III-B, Eq. (11); Sec. IV-L"},{"comment":"There is a one-sample label-alignment inconsistency between training and evaluation. Training uses features at time t to advance the state toward q_{t+1}, with force, gate, and condition labels aligned at t+1; the offline evaluators instead compare the time-t feature output with the force label time-stamped at t. At the platform sampling rate this is roughly a 17 ms shift, which can materially affect force metrics during contact onsets and releases. Please report force results under a consistent alignment (either train and evaluate at the same offset, or justify that the one-sample shift is negligible for the reported tasks).","section":"Appendix B"},{"comment":"For the payload benchmarks, the ground-truth force label is the nominal gravity vector [0,0,−mg]^T, assigned even while the payload is moving. The paper acknowledges this neglects object inertia and gripping friction. The very low force MAEs (0.08–0.12 N) may therefore reflect the model learning the task-conditioned constant label from telemetry and task identity rather than estimating the actual instantaneous external wrench. Please validate on the force-sensor dataset with the same metric as the payload benchmark, and/or add dynamic payload motions with measured force labels, to show that the payload numbers are not an artifact of a constant reference.","section":"Sec. III-C, Sec. IV-A"},{"comment":"The Franka benchmark uses a smooth commanded-pose proxy q˜_t^cmd = 1.03 q_rec_{t+5} constructed from a future recorded state. As the paper notes, these results are a future-state-conditioned offline benchmark, not an online evaluation. The conclusion and abstract nevertheless list Franka among the validation platforms for sensorless force estimation. Please either add a causal Franka evaluation (using the controller's actual commanded pose at time t) or explicitly scope the Franka claim in the abstract and conclusion to the offline, future-state-conditioned setting.","section":"Sec. IV-F and Appendix B"}],"minor_comments":[{"comment":"The caption contains incomplete sentences: “The shows the predicted magnitude” and “the shows the force-gauge ground truth” are missing legend identifiers. Please fix the caption.","section":"Fig. 1 caption"},{"comment":"The motor-condition comparison uses 32 training trajectories for NeuralActuator versus 16 for the handcrafted baselines, and evaluates at different decision granularities (framewise vs. windowed). This is disclosed but should be stated more prominently in the table or text to avoid an apples-to-oranges comparison.","section":"Sec. IV-G and Tab. XI"},{"comment":"The abstract reports “force MAE 0.47–0.73 N on SO-101” without stating that this is at a 500-step recorded-telemetry-conditioned rollout. Please include the horizon and evaluation condition in the abstract or refer to the table.","section":"Abstract and Sec. IV-F"},{"comment":"Training is described as running “up to 100000 epochs” with best checkpoints within 30000 epochs. This is likely iterations or training steps rather than epochs; please clarify the terminology.","section":"Sec. III-D"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern largely lands, but not as a circularity problem: the force head is independently supervised and there is no equation reducing predicted force to a fitted value. The real issues are identifiability and generalization. The implicit coupling allows the torque surrogate to absorb interaction effects, so the force head is not dynamically constrained; the payload labels are nominal during motion; and the label-alignment mismatch in Appendix B affects the reported force metrics. These are fixable within the manuscript's scope: a causal force evaluation on dynamic measured-force contacts, consistent label alignment, and a clearer scoping of the Franka benchmark would address the central concern. The dataset and code release are significant assets, and the paper is unusually transparent about its limitations."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"NeuralActuator is worth engaging with. It is the first system I know of that trains a torque surrogate through differentiable simulation from poses alone on low-cost servo arms, and it bundles that with a contact-gated force head and a released multi-platform dataset (NAD). The empirical work is unusually transparent: the paper states up front that the deployed implicit coupling leaves the force head out of the dynamics, that offline rollouts are telemetry-conditioned, and that the Franka benchmark is future-state-conditioned. Those are real limitations but they are disclosed, and the ablations on torque parameterization, force coupling, and architecture are the right checks to run. The force MAEs on payloads (0.08–0.12 N on OpenManipulator-X) are impressive if they hold up under independent test sets.\n\nThe soft spots are in proportion. The stress-test concern is legitimate. Under implicit coupling, tau_pred alone advances the simulation and f_hat is supervised but never applied, so the force head is not constrained by the rigid-body residual. The paper concedes the torque surrogate may absorb interaction effects. That means the strong force numbers could reflect correlations between motor current/task identity and the nominal gravity labels, rather than a physically grounded estimate. The \"unseen contact geometry\" checks are stationary holds, and the Franka benchmark is explicitly future-state-conditioned. So the paper does not establish causal generalization outside the labeled task set. That is not a fatal flaw: sensorless force estimation as a learned telemetry-to-force map is still useful, and the claim \"without F/T sensors at inference time\" is true. But the title-level promise of a physically coupled actuator model plus force perception is only partially met.\n\nA few smaller issues: training and evaluation use different label alignment (App. B, one-sample offset); motor-condition baselines get half the training data; force baselines omit a simple learned current-to-force regressor, which would be the natural lower bound; and some error tables lack variance. The label alignment mismatch is the most concerning because it can mask small systematic errors. None of these are load-bearing.\n\nWho is this for? People building sim-to-real pipelines for low-cost manipulators and anyone needing force feedback on servo arms without F/T sensors. It deserves a serious referee and probably a revision. I would send it out.","headline":"Solid, transparent empirical paper; the force head is a telemetry-to-force regressor in the deployed mode, which the paper discloses but which tempers the headline claim.","tokens_in":33247,"tokens_out":3394,"would_cite":true,"duration_ms":30443,"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":"NeuralActuator jointly learns a torque surrogate and sensorless force perception for low-cost servo robots, trained from pose trajectories alone.","keywords":["neural actuator model","torque surrogate","differentiable simulation","sensorless force estimation","contact probability gate","motor condition estimation","behavior cloning","low-cost servo robots"],"falsifier":"Mount a joint-torque sensor on the same arm and compare measured physical torque with the learned torque surrogate under matched commands and loads. If the surrogate differs from the measured torque by much more than the force-head error budget while pose prediction remains accurate, the surrogate is a kinematic fit; equivalently, freeze the shared features and retrain only the force head on a 600 g payload not present in training — a large error jump would show the force head has no independent physical grounding.","tokens_in":32356,"feed_emoji":"🤖","tokens_out":5430,"duration_ms":52662,"temperature":0.7,"pith_summary":"This paper tries to establish that low-cost servo-driven robots, whose actuators violate the linear current-to-torque assumption, can still be modeled accurately enough to serve as both dynamics simulators and force sensors. The proposed method, NeuralActuator, is a Transformer that reads a short history of commands, states, and actuator telemetry and outputs a torque surrogate, a contact-gated external force, and a motor-condition score. The torque surrogate is trained through differentiable simulation using only pose trajectories, removing the need for joint-torque sensors or calibrated current-to-torque maps. The paper reports payload force errors as low as 0.08–0.12 N on a 5-DoF arm and shows that feeding the predicted force to a behavior-cloning controller raises pick-and-place success from 80% to 92.5%. If correct, this would make sensorless force perception and force-aware control accessible to budget robot platforms.","feed_headline":"NeuralActuator estimates contact force on cheap servo arms to ~0.1 N","feed_subtitle":"Pose-only supervision replaces torque sensors; the learned force feedback lifts pick-and-place success to 92.5%.","key_machinery":"The load-bearing object is the torque surrogate τ_pred = g_θ(X_t), produced by a Transformer encoder over a nine-token history, which replaces the linear τ = K_t I assumption as the generalized control input to a differentiable rigid-body simulator. Training uses backpropagation through time on a pose Huber loss, so no torque sensing is needed. Alongside it, a force head outputs a raw 3D force multiplied by a contact-probability gate (f̂ = g·f̂_raw), and a condition head outputs a per-motor score. The design uses 'implicit coupling': the torque surrogate alone drives the simulator, while the force head is supervised in parallel but not applied as a generalized load — a choice the paper tests","core_discovery":"The paper's central claim is that on low-cost servo robots the standard linear current-to-torque law is unreliable, and a single Transformer-based model — fed a nine-frame history of commanded targets, proprioception, and actuator telemetry — can jointly predict (i) a torque surrogate that, when injected into a differentiable rigid-body simulator, reproduces real pose trajectories; (ii) a contact-gated external end-effector force; and (iii) a per-motor condition score. The torque surrogate is trained without any ground-truth joint-torque label: the only supervision is measured pose, compared with simulated pose after forward integration, and gradients flow back through the differentiable dyn","pith_inferences":["The force head may be largely mediated by the torque surrogate's ability to absorb external-load effects implicitly; dynamic or compliant contacts could expose this dependency, since the surrogate is only trained to match pose.","The pose-supervised surrogate scheme could transfer to other low-cost actuator families beyond the two servo types and one industrial arm tested, for example legged robots or dexterous hands with similar non-ideal actuators.","The released dataset and protocol could serve as a benchmark for actuator modeling, with the 'no torque labels' trick making large-scale teleoperation data collection feasible without instrumented joints."],"forward_implications":["Low-cost servo arms can get sensorless end-effector force estimation (order 0.1 N on payloads), enabling force-aware manipulation without force/torque sensors.","Differentiable simulation with pose supervision can substitute for torque sensing when calibrating actuator models.","Force feedback improves behavior-cloning success rates (80→92.5% pick-and-place, 85→95% lift-and-hold) and reduces over-current trips.","A few-trajectory fine-tune (12 trajectories, about 6 minutes) adapts the model to new hardware instances.","Motor-condition monitoring emerges from the same telemetry, achieving 91% accuracy and 0.95 AUC under controlled mechanical restriction."],"fun_headline_variants":["Transformer models servo torque and contact force without torque sensors","Sim-to-real torque and force from pose only on cheap robot arms","One neural net predicts torque, force, and motor health from telemetry","Pose-only supervision learns actuator dynamics and contact force","NeuralActuator: force and torque from pose trajectories, no torque sensors"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The force estimation only stays physically meaningful if the torque surrogate alone can explain the observed motion; the paper acknowledges that under implicit coupling the surrogate 'may absorb interaction effects needed to reproduce the observed motion,' which would make the force head a telemetry-to-force regressor rather than a physically grounded estimator.","fun_headline_variants_meta":{"raw":{"variants":["Transformer models servo torque and contact force without torque sensors","Sim-to-real torque and force from pose only on cheap robot arms","One neural net predicts torque, force, and motor health from telemetry","Pose-only supervision learns actuator dynamics and contact force","NeuralActuator: force and torque from pose trajectories, no torque sensors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000212,"raw_usage":{"total_tokens":1316,"prompt_tokens":870,"completion_tokens":446,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":614,"completion_tokens_details":{"reasoning_tokens":357}},"tokens_in":614,"tokens_out":446,"duration_ms":4269,"temperature":1.0,"reasoning_tokens":357,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T06:47:31.115054+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Mount a joint-torque sensor on the same arm and compare measured physical torque with the learned torque surrogate under matched commands and loads. If the surrogate differs from the measured torque by much more than the force-head error budget while pose prediction remains accurate, the surrogate is a kinematic fit; equivalently, freeze the shared features and retrain only the force head on a 600 g payload not present in training — a large error jump would show the force head has no independent physical grounding.","supporting_citations":[],"review_version":2}