{"id":"77d3db4d-de54-4370-8ff4-8b430fcba055","arxiv_id":"2608.05725","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A streaming spectral Poisson solver on an FPGA reconstructs visuotactile depth maps with deterministic 0.211 ms latency, enabling a 28 ms protective reflex loop.","lead":"This paper reports an FPGA pipeline that reconstructs 3D contact geometry from a visuotactile sensor in a fixed 0.211 ms with estimated power under 350 mW. A protective robot reflex closes in 28 ms, about six times faster than a host-based comparison.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 0.17% accuracy figure is a circular, affine-aligned comparison to the same software solver; absolute geometric accuracy is not validated.","rationale":"The reader's conditional verdict is appropriate. The fixed-latency result, 35,107 cycles or 0.211 ms, is measured on-chip over 1,000 frames and is not weakened by any concern raised here; the measured throughput of 10,106 frames per second is likewise a concrete hardware result. The power numbers are explicitly labelled simulation estimates, so they are not hidden overclaims. The reflex-loop comparison uses a different sensor in the host baseline, which weakens the attribution of the sixfold speedup but does not invalidate the reported 28.3 ms total. The main soft spot is the accuracy evaluation, and I agree with the reader that it is circular. I would sharpen the concern: the per-frame scale and offset alignment removes absolute calibration error, so the 0.17% figure is even narrower than a plain circular comparison. This affects the breadth of the accuracy claim rather than the core deterministic-latency contribution, so the verdict should remain conditional.","tokens_in":12235,"tokens_out":6556,"duration_ms":60483,"concrete_test":"Select at least three of the 15 tested contacts with independently known geometry, such as a machined cylinder, a sphere, and a flat punch of precisely measured depth. Reconstruct those contacts on-chip and compare the unaligned depth maps with physical ground truth obtained by a laser profilometer or from the calibrated indenter dimensions. Report the RMS error in millimetres and as a percentage of peak depth, and also report the fitted scale and offset values from the original alignment. If the unaligned error remains near 0.17% of peak depth and the fitted scale is close to 1, the physical accuracy claim is supported. If the unaligned error is much larger, the advertised accuracy should be restated as fixed-point fidelity to a software reference, with absolute depth accuracy left unverified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The accuracy claim is the least secure part of the central result. Methods, 'Reconstruction accuracy', compares the hardware output with a double-precision software implementation of the same spectral Poisson solver running on the same raw frames. The reported 0.17% RMS difference is therefore a fixed-point fidelity measure, not evidence about physical depth accuracy. The comparison is further weakened by a per-frame scale factor and offset that are fitted to the software reconstruction before the error is computed. This removes any global calibration-gain or offset error, so the number does not bound absolute depth error in millimetres. The Discussion states that 'the sensor outputs calibrated depth in millimetres', but the evaluation does not test that claim. If the photometric calibration, the homogeneous Dirichlet boundary condition, or the linear Poisson model is physically inaccurate for real contacts, the hardware inherits that bias and the advertised 'accurate geometry' claim is unsupported. The deterministic-latency and throughput measurements remain convincing, but the overall conclusion is only as strong as this unvalidated physical-model assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents a near-sensor computing architecture for visuotactile sensors, mapping a spectral Poisson solver onto a fully streaming FPGA datapath. The authors report a measured reconstruction latency of 35,107 cycles (0.211 ms at 166 MHz) across 1,000 consecutive frames, a throughput of about 10,106 frames/s, power estimates of 347 mW (FPGA cores) and 324 mW (45 nm ASIC), a 0.17% RMS difference from a double-precision software reference across 15 contact geometries, and a contact-to-motion reflex loop of 28.3 ms versus 169.9 ms for a host-based baseline. The central technical claim is that moving the Poisson solve onto the sensor node yields deterministic sub-millisecond latency and enables faster robotic reflexes. The paper includes detailed methods for latency characterization, accuracy evaluation, ASIC synthesis, and reflex-loop timing.","tokens_in":12416,"tokens_out":5795,"duration_ms":48536,"significance":"The paper's main contribution is a credible streaming hardware implementation of a direct spectral Poisson solver with deterministic, on-chip measured latency and throughput near the theoretical limit. The latency measurements are machine-checked over 1,000 frames, and the resource-scaling analysis is informative and reproducible. If the accuracy and system-level claims were properly validated, the architecture would be a significant step for high-speed visuotactile perception. However, the current evidence does not support the advertised absolute geometric accuracy or an unambiguous 'six times faster' reflex comparison; these gaps require revision before the paper can be accepted.","major_comments":[{"comment":"The 0.17% RMS difference is computed after a per-frame scale and offset are fitted to the double-precision software reference, which implements the same spectral Poisson algorithm (Supplementary S1). This metric therefore validates fixed-point arithmetic fidelity, not physical depth accuracy; it does not bound absolute error in millimetres. The Discussion's claim that 'the sensor outputs calibrated depth in millimetres' is unsupported because no comparison to known ground-truth geometry (e.g., indenters with independently measured dimensions) is provided. Please either add such a validation or revise the abstract and Discussion to describe the result as implementation fidelity rather than geometric accuracy.","section":"Methods, 'Reconstruction accuracy' (Fig. 3c); Discussion"},{"comment":"The near-sensor pathway uses the custom VBTS sensor, while the host baseline uses a GelSight Mini, a different sensor with a different frame rate and interface. The reported 6.0-fold contact-to-motion reduction therefore conflates sensor differences with the processing-location difference. To support the claim that near-sensor processing alone is responsible for the speedup, the comparison should use the same sensor with host-based reconstruction, or the latency contributions should be decomposed and reported separately (sensor capture, transfer, reconstruction, decision). As written, 'equivalent host-based loop using the same actuator' is misleading because the sensors differ.","section":"Methods, 'Reflex loop measurement' (Fig. 4c)"},{"comment":"The power comparisons mix estimation and measurement. The 347 mW FPGA and 324 mW ASIC values are simulation/synthesis estimates for the reconstruction cores, whereas the CPU, GPU, and Jetson values are measured device or package powers. The text discloses the measurement boundaries, but the abstract's 'energy-efficient' claim rests on an apples-to-oranges comparison. Please state explicitly that the near-sensor core power is an estimate, provide an end-to-end measured power figure for the prototype (including sensor and I/O), or temper the efficiency claim accordingly.","section":"Results, 'Deterministic latency and power efficiency' (Fig. 2e,f)"}],"minor_comments":[{"comment":"The figure labels use 'float64' while the text uses 'double-precision'; please unify the terminology.","section":"Fig. 3 and text"},{"comment":"The text states 'no full-frame input buffering' but also describes the two transform passes as connected by 'full-frame matrix transpositions'; please clarify that the input is streamed while intermediate transposition buffers are full-frame.","section":"Methods, 'Reconstruction pipeline'"},{"comment":"The abstract reports 347 mW, while Fig. 1b also lists a 324 mW ASIC estimate; please state which process technology the headline power figure refers to.","section":"Abstract and Fig. 1b"},{"comment":"The measured cycle counts scale by about 3.7x per grid-size doubling (9,421 to 35,107 to 135,448), described as 'approximately fourfold'; please report the expected O(N^2) values or explain the small deviation.","section":"Results, 'Deterministic latency and power efficiency' (Fig. 2c)"},{"comment":"Reference [28] is a patent application; if it is the sole source for the underlying sensor framework, please add publicly available citations for the sensor design and calibration.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The deterministic latency and throughput results are solid and are the paper's core strength. The main risks are overclaiming absolute geometric accuracy (the accuracy measure is a fixed-point fidelity check against the same algorithm, with per-frame alignment) and the confounded reflex-loop comparison that uses different sensors. A revision that adds absolute ground-truth validation and a same-sensor host-baseline comparison would make the paper publishable; without these, the headline claims are stronger than the evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: this paper has a real, measured hardware result — deterministic sub-millisecond Poisson reconstruction on an FPGA for a visuotactile sensor — and it's worth taking seriously. The 0.17% accuracy number in the abstract, however, is not a validation of geometric accuracy. It's a fixed-point fidelity check against the same algorithm in software, and the stress-test note got that right. What's genuinely new: the fully streaming spectral Poisson pipeline with a fixed cycle count, measured on-chip at 35,107 cycles (0.211 ms) for 128x128 frames. That's a clean engineering result, backed by an honest latency distribution plot and a throughput measurement within 0.3% of the theoretical limit. The reflex loop is also a nice demonstration: the 28 ms contact-to-motion latency is real, and the comparison with a host-based loop on the same actuator is a fair way to isolate the sensing-and-decision part, even though the two pathways use different sensors. The paper is also transparent about the concurrent FasTac work, and the resource-scaling analysis is sensible. The soft spots are in the accuracy and power claims. The 0.17% RMS difference is computed against a double-precision software reconstruction of the same raw frames, with a per-frame scale and offset fitted before comparison. That removes any global calibration error, so the number bounds only the fixed-point arithmetic error, not the physical depth error in millimetres. The paper's stated claim that 'the sensor outputs calibrated depth in millimetres' is not supported by the evaluation. The power figures are simulation estimates from Vivado and Design Compiler — useful for a prototype but not measured. The reflex speedup is also confounded by the sensor difference: the near-sensor path uses your custom VBTS, while the baseline uses a GelSight Mini, so the 6x improvement is not purely from the near-sensor processing. None of this breaks the central latency and throughput contribution, but the abstract's 'accurate' should be tempered or backed with a real ground-truth contact geometry test (e.g., machined indenters with known depths). Who should read this: anyone working on visuotactile sensors, near-sensor computing, or hardware acceleration for robotics. The paper deserves a serious referee — the engineering is solid and the latency result is impressive. I'd recommend a conditional accept, with requests to fix the accuracy language and ideally add a physical depth validation. The math and citation pattern look sound, and the authors are not overselling their contribution relative to prior work. Bring it to the reading group; it will spark a useful debate about what counts as validation in hardware papers.","headline":"Solid, measured deterministic-latency hardware with an over-stated accuracy claim; the 0.17% figure is a fixed-point fidelity check, not a geometric accuracy validation.","tokens_in":783,"tokens_out":886,"would_cite":true,"duration_ms":24131,"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":"This paper claims that a spectral Poisson solver, implemented as a fully streaming FPGA pipeline, reconstructs dense tactile depth maps with a fixed 0.211 ms latency, 0.17% error versus a double-precision reference, and triggers a…","keywords":["visuotactile sensing","near-sensor computing","spectral Poisson solver","FPGA pipeline","deterministic latency","fixed-point arithmetic","robotic reflex"],"falsifier":"Take one of the 15 test indenters and scan its contact face with a calibrated laser profilometer to obtain physical ground truth, then compare the hardware depth maps against that scan; if the RMS deviation from the physical surface is much larger than 0.17% of peak depth, the hardware's advertised accuracy claim is unsupported.","tokens_in":12063,"feed_emoji":"⚡","tokens_out":5205,"duration_ms":41212,"temperature":0.7,"pith_summary":"This paper tries to establish that dense tactile geometry can be reconstructed at the sensor itself, fast enough and cheaply enough for robotic reflexes. The key move is implementing the standard spectral Poisson solver, which recovers depth from measured surface gradients, as a fully streaming hardware pipeline with no iterative loops and no data-dependent branching. On an FPGA running at 166 MHz it produces the first depth value of each 128x128 frame in a fixed 0.211 ms, reconstructs 15 contact geometries within 0.17% of a double-precision reference, and closes a protective reflex loop in 28.3 ms versus 169.9 ms for a host-based pipeline. The claim matters because visuotactile sensors have rich spatial resolution but have been limited by host-side reconstruction latency and jitter.","feed_headline":"Tactile depth in a fixed 0.211 ms","feed_subtitle":"A streaming FPGA solver reconstructs contact geometry at the sensor, powering reflexes six times faster than a host loop.","key_machinery":"The load-bearing object is the spectral Poisson solver: the observation that the discrete Laplacian is diagonalized by the discrete sine transform, turning the Poisson equation into $N^{2}$ independent scalar divisions in the frequency domain. In hardware this solver becomes a fixed-schedule streaming pipeline: gradient lookup, finite differencing, forward DST, point-wise spectral division, and inverse DST, with the 2D transform computed separably and each 1D transform implemented by a pipelined FFT core with input reordering and phase correction. This is what makes latency deterministic by construction rather than by measurement.","core_discovery":"The central discovery is a complete mapping of the depth-reconstruction Poisson equation onto a single feed-forward hardware datapath. The solver diagonalizes the discrete Laplacian under homogeneous Dirichlet boundary conditions, so the 2D Poisson solve reduces to a forward discrete sine transform, an element-wise division in the spectral domain, and an inverse transform; the hardware implements these as pipelined streaming cores with 24-bit fixed-point arithmetic. Because the pipeline has no iteration or input-dependent control flow, every frame takes exactly 35,107 clock cycles at 128x128, which is 0.211 ms at 166 MHz, and the same datapath runs at 64x64 and 256x256 with quadratic latency scaling. The reconstructed depth maps agree with a double-precision software reference to 0.17% RMS of peak contact depth over 15 geometries, and an on-chip threshold comparator acting on the depth stream triggers finger withdrawal in 28.3±4.9 ms, about six times faster than a host-based loop using the same actuator.","pith_inferences":["If the underlying spectral Poisson model is physically accurate for real contacts, the same architecture could be reused for any sensing modality that produces gradient fields and needs fast depth, not just visuotactile skins.","The paper's accuracy comparison validates fixed-point arithmetic against floating-point software, not against physical ground truth; a profilometer-based validation would be the natural next test.","Event-based tactile sensors and this dense-depth pipeline are complementary: events flag fast changes while this pipeline supplies continuous geometry, so fusing them could cover both slip transients and steady contact shape.","Because only on-chip memory limits grid size, adding external memory could push the deterministic pipeline beyond 256x256 without changing the solver core."],"forward_implications":["Every 128x128 tactile frame can be reconstructed to depth in a fixed 0.211 ms, making tactile geometry available at the camera frame rate with no frame-to-frame jitter.","A threshold comparator fed directly by the depth stream can trigger a protective reflex in 28.3 ms, below the reported human withdrawal-reflex range of 65–137 ms.","The same streaming architecture scales to 64x64 and 256x256 with latency proportional to pixel count, so higher-resolution tactile depth remains deterministic.","At 324 mW estimated in a 45 nm ASIC implementation, the reconstruction datapath is small and power-efficient enough to consider integrating into a fingertip-scale sensor."],"supporting_citations":[{"why":"supplies the spectral diagonalisation of the discrete Laplacian that the solver implements.","marker":"[31]"},{"why":"defines the visuotactile reconstruction task by recovering geometry from surface gradients via Poisson integration.","marker":"[1]"},{"why":"provides the fast cosine/sine transform method used for the streaming 1D DST cores.","marker":"[32]"},{"why":"is the sensor design lineage on which the custom visuotactile sensor is based.","marker":"[9]"},{"why":"motivates the near-sensor computing approach that this architecture carries into tactile sensing.","marker":"[19]"},{"why":"represents the iterative hardware PDE solvers whose variable latency this pipeline is contrasted against.","marker":"[23]"},{"why":"supplies the human nociceptive withdrawal reflex latency range used to frame the reflex result.","marker":"[29]"},{"why":"is the concurrent DST-based visuotactile FPGA solver used as a comparison point for latency and throughput.","marker":"[34]"}],"fun_headline_variants":["Streaming Poisson solver fixes tactile depth at 0.211 ms","Near-sensor depth: deterministic 0.211 ms per frame","Tactile reflexes six times faster via near-sensor solver","Fixed-latency tactile depth: 0.211 ms, no host needed","On-chip Poisson solve delivers tactile depth in 0.211 ms"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The accuracy and reflex claims stand on treating the double-precision software Poisson reconstruction as the true contact geometry; if that spectral model misrepresents real contacts, the hardware inherits the error.","fun_headline_variants_meta":{"raw":{"variants":["Streaming Poisson solver fixes tactile depth at 0.211 ms","Near-sensor depth: deterministic 0.211 ms per frame","Tactile reflexes six times faster via near-sensor solver","Fixed-latency tactile depth: 0.211 ms, no host needed","On-chip Poisson solve delivers tactile depth in 0.211 ms"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000432,"raw_usage":{"total_tokens":2218,"prompt_tokens":977,"completion_tokens":1241,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":593,"completion_tokens_details":{"reasoning_tokens":1148}},"tokens_in":593,"tokens_out":1241,"duration_ms":8637,"temperature":1.0,"reasoning_tokens":1148,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T14:35:18.506667+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take one of the 15 test indenters and scan its contact face with a calibrated laser profilometer to obtain physical ground truth, then compare the hardware depth maps against that scan; if the RMS deviation from the physical surface is much larger than 0.17% of peak depth, the hardware's advertised accuracy claim is unsupported.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the spectral diagonalisation of the discrete Laplacian that the solver implements."},{"cited_title":"& Adelson, E","cited_arxiv_id":null,"evidence_quote":"defines the visuotactile reconstruction task by recovering geometry from surface gradients via Poisson integration."},{"cited_title":"A fast cosine transform in one and two dimensions.IEEE Trans","cited_arxiv_id":null,"evidence_quote":"provides the fast cosine/sine transform method used for the streaming 1D DST cores."},{"cited_title":"In2014 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 3988–3993 (2014)","cited_arxiv_id":null,"evidence_quote":"is the sensor design lineage on which the custom visuotactile sensor is based."},{"cited_title":"& Chai, Y","cited_arxiv_id":null,"evidence_quote":"motivates the near-sensor computing approach that this architecture carries into tactile sensing."},{"cited_title":"R., Reguly, I","cited_arxiv_id":null,"evidence_quote":"represents the iterative hardware PDE solvers whose variable latency this pipeline is contrasted against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the human nociceptive withdrawal reflex latency range used to frame the reflex result."},{"cited_title":"FasTac: A Curved Multispectral Vision-Based Tactile Sensor for High-Speed High-Precision 3D Shape and Force Perception","cited_arxiv_id":"2607.28416","evidence_quote":"is the concurrent DST-based visuotactile FPGA solver used as a comparison point for latency and throughput."}],"review_version":2}