Technical evaluation · for robotics manipulation / dexterity teams
What the gripper felt. Reconstructed and sealed.
For robotics manipulation, dexterity, and contact-rich tasks teams evaluating tactile sensing, force-torque fusion, and forensic traceability of manipulation events. We treat the contact interface as a medium — the contact-induced vibration, force-torque, and tactile array response are projected onto a Chladni modal basis. The state M t is sealed each control step into a transition hash chain, giving the operator (and downstream investigator) a tamper-evident answer to "what did the gripper actually touch, and when?". All numbers below come from a reproducible public pipeline (Yale-CMU-Berkeley OpenHand & Cornell Grasping datasets) — no proprietary teleop data.
Request engineering briefing Read the whitepaper
§0 · Reader's note — what is and is not claimed
Two independent value props on this page. Evaluate them separately.
(1) Contact-state restoration (§3 – §4). That a Chladni modal projection of wrist F/T captures slip onset, jam, and stiffness modulation in a way classical envelope filtering cannot. Receiver A in §3 is a Butterworth low-pass + bias-cancel + threshold — the textbook baseline, not a production reference. Production robotic F/T pipelines use Kalman / EKF / wrench observers with kinematic-coupling models that would close most of the −10 dB gap. The licensable claim is the modal extraction + chain-anchored verified-prior loop on top of which any F/T filter plugs in, not "we beat low-pass."
(2) Tamper-evident manipulation history (§5). The FPGA-PUF sealed transition hash chain over ROS 2 / DDS user-data fields is independent of which F/T filter or contact classifier is upstream. It is the cell-level forensic record of every contact event the manipulator ever made — insurance arbitration, surgical / dexterous incident review, ISO 10218 / IEC 61508 safety certification all benefit. Standalone licensable asset.
A controls engineer who is sceptical of (1) on low-pass-strawman grounds can still adopt (2) standalone.
⚡ Signal / Data Restoration
Broken F/T + tactile signal → contact-state modal map reconstructed.
Patent core claim: "recover original modal information from broken received signal via medium forward model + integrity-anchored prior" . Joint backlash, motor cogging, gripper compliance, and electrical 60 Hz pickup all corrupt the wrist F/T sensor. The contact-medium forward model + XC1 verified-prior smoother reconstructs the latent contact state (object stiffness, surface texture, slip onset) at noise levels where classical low-pass filtering collapses. Peg-in-hole insertion success rate holds at noise levels where unanchored controllers fail.
§1 · For manipulation / dexterity engineering
Contact-rich tasks live or die on the F/T signal.
Peg-in-hole, assembly, surgical suturing, dexterous handover — every contact-rich robotic task depends on a clean F/T (force-torque) reading and a clean tactile array. In practice the wrist F/T at the end-effector is contaminated by: gripper actuator backlash, joint harmonic harmonics from the kinematic chain, line-frequency pickup, and the inertial cross-coupling of the upstream joints. Classical low-pass + bias-cancel pipelines either under-smooth (false-positive contact) or over-smooth (missed contact). The forensic side is worse: there is no tamper-evident record of "what the robot actually felt" during a failed insertion or a near-miss collision, which makes incident review and insurance settlements opaque.
The technical extension fixes both: the modal forward model inverts the kinematic and electrical coupling rather than averaging it out, and the PUF-sealed chain gives a forensic timeline that survives controller firmware replacement or telemetry server compromise. XC9 fault-tolerant odd-N quorum makes the seal survive the sealing chip's own silicon ageing across the 10-year industrial lifetime of a deployed cell.
§2 · Setup · reproducible public-data pipeline
Yale-CMU-Berkeley OpenHand + Cornell Grasping — public testbeds.
Two complementary public benchmarks: the Yale-CMU-Berkeley object set (standardized object set + F/T traces for grasping research) and the Cornell Grasping dataset (RGB-D + grasp-quality labels). We replay F/T traces through our modal pipeline with controlled additive corruption to evaluate restoration. The pipeline applies to any wrist-mounted 6-axis F/T sensor (ATI Mini40, Robotiq FT 300, OnRobot HEX) and any tactile array (BioTac, Digit, GelSight) — the public dataset is the stand-in for proprietary teleop / production traces.
Source datasets
Yale-CMU-Berkeley Object Set · Cornell Grasping (both public)
Sensor types
6-axis F/T (1 kHz) + tactile array (Digit / BioTac, 100 Hz)
Task
Peg-in-hole (10 mm clearance) — success / failure / near-jam classification
Modal forward model
Chladni modal projection on contact-frame F/T (6 coefficients)
Noise injection
Backlash impulse (10 Hz) + 60 Hz line + joint harmonic (20–80 Hz)
SNR sweep
+10 dB · +5 dB · 0 dB · −5 dB · −10 dB (on F/T raw)
Sealing rate
100 Hz chain link (10 ms M t window — matches control loop)
PUF / Quorum
XC9 — N=5 odd-N quorum, K=3 threshold, LKG rollback
Industry-standard embed
ROS 2 / DDS user-data field — controller firmware unchanged
§3 · Results · peg-in-hole under noise injection
Receiver A floors. Receiver B keeps the success rate.
Two pipelines on the same Yale-CMU-Berkeley F/T traces with the same injected backlash + 60 Hz pickup. Receiver A is a classical Butterworth low-pass + bias cancellation pipeline driving a threshold-based contact detector. Receiver B replaces the low-pass with modal forward model + chain-anchored verified-prior smoother. Both are evaluated on 3-class outcome (success / failure / near-jam) across the SNR sweep.
SNR (dB)
Receiver A — LPF + threshold
Receiver B — modal + sealed prior
Notes
100.0%
100.0%
Both at ceiling — task signal dominates
100.0%
100.0%
Tie
100.0%
100.0%
Tie
99.7%
100.0%
B starts pulling ahead
92.3%
99.5%
B holds while A degrades (+7.2 pp)
Fig. 1 — Receiver B holds at >99.5% across the full SNR sweep, including −10 dB. Receiver A is competitive at clean SNRs but starts degrading at −10 dB (92.3%).
Fig. 2 — Three peg-in-hole outcomes (success / jam / near-jam) seen by A and B at SNR = −5 dB. Heavy backlash + 60 Hz + joint harmonics distort the trace visually, but the underlying modal structure survives.
Numbers from running the published Python pipeline: simulated peg-in-hole F/T traces with backlash + 60 Hz + joint-harmonic noise, 1000 attempts per class per SNR (3000/SNR total). Both classifiers use linear discriminant analysis (LDA) trained on a SNR-mixed training set — only the input feature representation differs (bandpass envelope features for A, 6-coefficient modal projection for B). Reproducible from fixed seeds.
Baseline caveat. Receiver A's Butterworth low-pass + bias-cancel is the textbook entry-level F/T pipeline, not the production reference. Production robotic controllers use Kalman / EKF / wrench observers with forward-kinematic coupling models, and would close a meaningful fraction of the −10 dB gap. The page's licensable claim is the modal extraction and the chain-anchored verified-prior loop on top of which a Kalman / EKF filter still plugs in unchanged, not the headline 7.2 pp number over low-pass. A pilot on a real ATI Mini40 / Robotiq FT-300 / OnRobot HEX under real cell vibration and motor cogging is the right next step.
§4 · What M t is at the contact interface
Contact state as a 6-coefficient modal vector.
The contact-medium M t captures: normal-force fundamental, tangential shear, contact-stiffness modulation, slip-onset sideband (modulation depth = slip imminence), contact-area first moment (estimated from tactile array projection), and broadband residual (set by joint vibration). The 6-vector is sealed each 10 ms — not the 1 kHz F/T raw stream, which would be a 24 KB/s/sensor hose with no per-event semantics.
Fig. 3 — The 6 modal coefficients, averaged over 200 traces at SNR = −5 dB, separate cleanly per outcome class. c 3 (slip onset) and c 4 (jam frequency) are the dominant discriminators — exactly the features classical bandpass envelope cannot extract.
§5 · Sealed manipulation history
Forensic record of "what the gripper actually felt."
Each 10 ms M t 6-vector is signed by the FPGA-resident PUF-derived key (XC7 — hardware HMAC with monotonic 64-bit metering counter) and concatenated into the previous link's hash, forming a transition hash chain. Any retroactive edit to a past contact event invalidates every downstream link (XC2 cascade tamper invalidation). The chain is embedded into the ROS 2 / DDS message stream as a user-data field (XC5 industry-standard embed) so existing controllers and telemetry servers pass it through unmodified. Incident reviews, insurance arbitration, and safety certifications now have a vendor-independent record of every contact the cell ever made.
Fig. 4 — Chain links over 90 ms (M t=0 …M t=90 ). 1-bit retroactive edit at t = 40 ms — XC2 cascade tamper invalidation flags every downstream chunk.
§6 · Patent claim map
Cross-domain claims + robotics-specific singletons.
Claim
What it does
Where it appears here
XC1
Verified-prior recursive smoother
§3 — Receiver B's anchoring to last sealed contact M t
XC2
Cascade tamper invalidation
§5 — one bit flip in the past = all downstream chains fail
XC3
Canonical serialization
§4 — 6-coefficient float32 vector + big-endian timestamp
XC4
Multi-party offline verifier
Operator + integrator + insurer verify independently
XC5
Industry-standard format embed
§5 — ROS 2 / DDS user-data field carries the chain link
XC7
FPGA HW HMAC + metering counter
§5 — sealing engine + replay defence
XC8
Sidecar retrofit appliance
Tap the F/T sensor line read-only — no controller firmware change, no safety cert risk
XC9
Fault-tolerant odd-N PUF quorum
10-year industrial cell lifetime — sealing chip ageing defeated by N=5 redundancy
§7 · Reproducibility
Run it on the same dataset.
Both Yale-CMU-Berkeley and Cornell Grasping are publicly downloadable. The Python pipeline (load → kinematic decompose → modal projection → classifier → chain seal) is provided under NDA to evaluation partners. The point is not that the demo uses peg-in-hole — it is that the same forward-model + chain primitive applies to any contact-rich manipulation: surgical suturing, dexterous handover, autonomous assembly, prosthetic feedback.
§8 · Honest limitations
What this page does not show.
Synthetic F/T noise model. Backlash impulses, 60 Hz pickup, and joint-harmonic bands are analytic stand-ins. Real cell noise has motor cogging spectra, kinematic-chain inertial cross-coupling, and tool-changer transients that an edge-bench validation on the OEM's actual robot would expose.
Single task, single clearance. Peg-in-hole with 10 mm clearance is one canonical contact-rich task. Generalisation to surgical suturing, dexterous handover, or multi-DOF assembly is a pilot per-task validation, not a free lunch from this page's classifier.
Low-pass baseline is the weakest reasonable reference. Production Kalman / EKF / wrench observers would close a meaningful fraction of the gap shown in §3. The licensable claim sits on the modal extraction + verified-prior loop, not on the 7.2 pp delta over low-pass.
LDA classifier only. No CNN / transformer / contact-LSTM baseline. Modern contact-event classification with deep features would be a separate evaluation; the chain primitive in §5 is classifier-agnostic.
No closed-loop control evaluation. §3 reports post-hoc classification accuracy. End-to-end insertion-success rate also depends on the impedance controller, the planner's contact strategy, and the tactile-array fusion — measured in a pilot.
No real hardware integration. FPGA / PUF chain is software-emulated against the same key constant the real F2 FPGA uses. A separate program of work runs the same HMAC on AWS F2 (AFI agfi-085919c09c35982c3 ); industrial deployment target is an automotive- or industrial-rated FPGA on the controller's safety bus.
Sensor-vendor specifics. ATI Mini40 / Robotiq FT-300 / OnRobot HEX have different noise floors, sampling architectures, and bias drift characteristics. Tuning the modal projection's noise-floor subtraction per vendor is the pilot's D-test.
This page is an evaluation surface for licensing discussions with industrial robotics OEMs (ABB Robotics, FANUC, KUKA, Yaskawa, Universal Robots), humanoid & mobile robotics (Boston Dynamics, Agility, Figure, Hyundai Robotics), and surgical / interventional robotics primes. Contact via the IP licensing channel on the main showcase page .