dsfb-robotics

DSFB Structural Semiotics Engine for Robotics Health Monitoring — a deterministic, no_std + no_alloc + zero-unsafe observer layer that reads the residual streams existing robot control and prognostics pipelines already compute, and structures them into a human-readable grammar of typed episodes.

Status: Version 1.0 (April 2026). The full crate ships: core DSFB pipeline (grammar FSM, envelope, engine, kinematics / balancing helpers), 20 dataset adapters, paper-lock binary, figure pipeline, Colab notebook, audit stack, and 81-page companion paper. Per-revision history in CHANGELOG.md; architectural notes in ARCHITECTURE.md.

What this crate is

A read-only observer that takes residuals already produced by incumbent robotics pipelines — joint torque identification residuals, inverse-dynamics residuals, whole-body MPC contact-force residuals, centroidal-momentum observer residuals, bearing envelope-spectrum residuals, health-index trajectories — and emits a three-state grammar (Admissible / Boundary / Violation) with typed episode records and provenance-tagged audit trails.

The upstream pipelines keep running unchanged. Removing DSFB changes nothing about the robot's control or safety behaviour.

What this crate is not

DSFB is not a competitor to any existing robotics method. It does not:

• Classify bearing faults or identify root cause
• Provide calibrated Pd/Pfa, F1, ROC-AUC, or confusion matrices
• Detect faults earlier than threshold alarms, CUSUM, EWMA, or RMS monitors
• Predict remaining useful life (RUL)
• Replace inverse-dynamics identification, Kalman/Luenberger observers, whole-body controllers, MPC, or any incumbent observer
• Guarantee hard real-time latency under any specific controller platform
• Provide ISO 10218-1/-2:2025, IEC 61508, or ISO 13849 certification

Existing methods continue to outperform DSFB at their own tasks. DSFB's role is orthogonal: it recovers structure from the residuals those methods discard.

Architectural contract

Canonical API

use dsfb_robotics::{Episode, observe};

let residuals: &[f64] = &[0.01, 0.02, 0.05, 0.12, 0.21];
let mut out = [Episode::empty(); 16];
let n = observe(residuals, &mut out);

for e in &out[..n] {
    // advisory only — no write-back, no upstream coupling
    let _ = (e.index, e.grammar, e.decision);
}

Episode fields are byte-identical to the canonical form in dsfb-semiconductor so downstream tooling consumes DSFB episodes uniformly across crates.

Feature flags

Dataset evaluation (companion paper, twenty real-world datasets)

The companion paper at paper/dsfb_robotics.tex evaluates DSFB on twenty public real-world datasets across three families. Every dataset is a physical-hardware recording under a permissive licence (Apache-2.0 / MIT / CC-BY-4.0 / CC-BY-SA-4.0 / BSD-3-Clause / academic-fair-use). Zero synthetic or simulated data is admitted.

Per-dataset provenance, SHA-256 checksums, and fetch instructions live at data/processed/PROCESSED_MANIFEST.json. All twenty processed-residual CSVs ship in-tree; raw upstream-source data is fetched on demand by scripts/preprocess_datasets.py.

Dataset honesty disclosure

• Real data only. No synthetic data is mixed with real results anywhere in this crate or its paper. The only micro-fixtures used in unit tests are clearly illustrative arrays of ≤10 values (e.g. [0.1, 0.2, 0.5, 1.2, 2.1]).
• No simulation frameworks. MuJoCo, Isaac, Gazebo, RaiSim, Drake, Webots, and PyBullet are not used anywhere in this crate.
• Paper-lock fallback policy. When invoked without the required real dataset at the documented path, paper-lock <slug> exits with a clear error pointing to the relevant oracle-protocol doc under docs/. It never silently substitutes a synthetic fixture.

Reproducibility

End-to-end reproduction recipe lives in REPRODUCE.md. One-command form:

cargo run --release --bin paper-lock --features std,paper_lock -- <slug>

A Colab notebook at colab/dsfb_robotics_reproduce.ipynb bundles the in-tree processed CSVs and reproduces every paper figure end-to-end. Free-tier Colab Run-All budget: ~30 min cold (LTO build dominates); ~4 min on a developer laptop with cached target/.

Audit

The crate ships a layered audit stack — every artefact is reproducible from the committed source and documented inline.

The audit/README.md indexes every artefact with reproduction commands.

Licensing

• Reference implementation: Apache-2.0 (see LICENSE).
• Theoretical framework and supervisory methods: proprietary Background IP of Invariant Forge LLC; commercial deployment requires a separate written licence. See NOTICE.
• Datasets: each dataset retains its upstream licence; see data/processed/PROCESSED_MANIFEST.json.

Licensing enquiries: licensing@invariantforge.net

Companion paper

paper/dsfb_robotics.tex — 81-page LaTeX specification of the DSFB framework applied to robotics health monitoring (Version 1.0, April 2026). The paper includes the augmentation-thesis hero figure (§1.4), the twenty-dataset evaluation (§10) with bootstrap CIs / sensitivity grid / ablation / Cohen's d effect size, the worked example on the Gaz 2019 Panda dataset (§11), an explicit Non-Claims table, a failure-modes section, a limitations section with the 50-point engineering-criticisms subsection, the Falsifiability statement, the Non-Intrusion Manifest appendix, and the Motif Gallery. Every empirical number in §10 is reproducible from this crate under paper-lock.

Citation

If you use this crate or reference its companion paper, please cite:

de Beer, R. (2026). DSFB Structural Semiotics Engine for Robotics Health Monitoring: A Deterministic Augmentation Layer for Typed Residual Interpretation of Joint Degradation, Actuator Drift, and Kinematic Anomalies in Safety-Critical Robotic Systems (v1.0). Zenodo. https://doi.org/10.5281/zenodo.19778382

BibTeX:

@software{debeer_2026_dsfb_robotics,
  author    = {de Beer, Riaan},
  title     = {DSFB Structural Semiotics Engine for Robotics Health Monitoring:
               A Deterministic Augmentation Layer for Typed Residual Interpretation
               of Joint Degradation, Actuator Drift, and Kinematic Anomalies in
               Safety-Critical Robotic Systems},
  version   = {v1.0},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.19778382},
  url       = {https://doi.org/10.5281/zenodo.19778382}
}

Machine-readable metadata: see CITATION.cff. Formal bibliographic entries are maintained in the companion paper's bibliography.

Authorship & co-authorship policy

This crate and its paper are authored by Riaan de Beer (Invariant Forge LLC).