Datasets:

kordelfrance
/

scentience-olfaction-skills

name: olfactory-navigation
description: >
  Interprets a time-ordered buffer of Scentience OPU sensor readings to classify
  plume contact state (approaching, turbulent contact, signal loss,
  near-source), then recommends a navigation action with calibrated confidence.
  Use when controlling a robot or UAV through an odor plume, recovering from
  signal loss, following a chemical gradient, or evaluating real-time olfactory
  source localization.
version: 1.0.0
author: kordelfrance
license: Apache 2.0
compatibility: >
  Works with raw JSON readings from the scentience BLE or Sockets API. No
  additional SDK install required for interpretation. Use the ble-device skill
  to acquire readings. Optional: scentience Navigation Environment
  (gym-compatible) for simulation and RL.
metadata:
  domain: olfaction
  tags:
    - navigation
    - plume-tracking
    - source-localization
    - robotics
    - uav
    - gradient
    - casting
  hardware:
    - Reconnaisscent
    - Scentinel
  depends_on:
    - ble-device
  related_papers:
    - 'Chasing Ghosts: A Sim2Real Olfactory Navigation Stack (arXiv 2026)'
    - >-
      Emergent Behavior in Evolutionary Swarms for Machine Olfaction (ACM GECCO
      2024)
    - 'AI and Olfaction: A Survey on the Sense of Smell for Robotics (ChemRxiv)'

Goal

Given a sliding window of Scentience OPU readings, classify the robot's current plume interaction state and emit a structured navigation recommendation — including action, confidence, and reasoning — suitable for direct consumption by a motor controller or higher-level planner.

Instructions

1. Buffer Readings

Accumulate a sliding window of N readings (N = 10–30 depending on polling rate). Extract the target compound. Use VOC as the default; substitute a specific compound (e.g., NH3, H2S) when tracking a known chemical source.

from collections import deque
window = deque(maxlen=30)

async for reading in device.stream_ble(async_=True):
    if reading["STATUS_opuA"] != "active":
        continue
    window.append(reading)
    if len(window) >= 10:
        result = analyze_plume(window, compound="VOC")
        controller.act(result)

2. Compute Signal Features

import numpy as np

def extract_features(window, compound: str, noise_floor: float = 0.02):
    signal = [r[compound] for r in window]

    # Baseline correction — rolling minimum over the window
    baseline = min(signal)
    normalized = [max(0.0, s - baseline) for s in signal]

    n = len(normalized)
    recent  = normalized[max(0, n - 3):]
    prior   = normalized[max(0, n - 10): max(0, n - 3)]

    short_trend  = np.mean(recent) / (np.mean(prior) + 1e-6)
    long_trend   = float(np.polyfit(range(n), normalized, 1)[0])
    intermittency = sum(1 for s in normalized if s > noise_floor) / n
    peak_snr     = (max(normalized) - np.mean(normalized)) / (np.std(normalized) + 1e-6)

    return {
        "short_trend": short_trend,
        "long_trend": long_trend,
        "intermittency": intermittency,
        "peak_snr": peak_snr,
        "mean_signal": float(np.mean(normalized)),
    }

Feature	Meaning
`short_trend`	Ratio of recent 3-reading mean to prior 7-reading mean
`long_trend`	Linear slope over the full window (positive = strengthening)
`intermittency`	Fraction of readings above detection threshold
`peak_snr`	Signal quality; values < 2 indicate noise-dominated readings

3. Classify Plume State

Apply this decision matrix (order matters — evaluate top-to-bottom, take the first match):

State	Trigger conditions	Physical interpretation
`NEAR_SOURCE`	`intermittency > 0.7` AND `short_trend < 1.15`	Saturating; likely within source region
`APPROACHING`	`short_trend > 1.2` AND `long_trend > 0`	Signal strengthening; moving toward source
`CONTACT_TURBULENT`	`intermittency > 0.4` AND `peak_snr > 2`	Inside plume; turbulent filament contact
`LOST_PLUME`	`intermittency < 0.15` (after prior contact)	Plume exit or plume shifted
`UNKNOWN`	None of the above	Insufficient or noisy signal

def classify_state(features, had_prior_contact: bool) -> str:
    f = features
    if f["intermittency"] > 0.7 and f["short_trend"] < 1.15:
        return "NEAR_SOURCE"
    if f["short_trend"] > 1.2 and f["long_trend"] > 0:
        return "APPROACHING"
    if f["intermittency"] > 0.4 and f["peak_snr"] > 2:
        return "CONTACT_TURBULENT"
    if f["intermittency"] < 0.15 and had_prior_contact:
        return "LOST_PLUME"
    return "UNKNOWN"

Apply state hysteresis: require 3 consecutive readings in a new state before transitioning. Single-frame transitions are treated as noise.

4. Map State to Navigation Action

State	`recommended_action`	Notes
`APPROACHING`	`"advance"`	Maintain heading; increase sampling rate
`CONTACT_TURBULENT`	`"hold_and_cast"`	Hold position; lateral sweeps to center on filament
`LOST_PLUME`	`"cast_upwind"`	Crosswind casting with upwind bias; expand search radius
`NEAR_SOURCE`	`"slow_and_densify"`	Reduce speed; increase spatial sampling density
`UNKNOWN`	`"hold"`	Wait for window to fill; do not act on ambiguous state

If wind vector is available (ENV_* fields or external anemometer), bias cast_upwind direction using atan2(wind_y, wind_x). Without wind data, cast orthogonally to the last known heading.

5. Format Output

{
  "timestamp_ms": 1743379215000,
  "compound": "VOC",
  "plume_contact_confidence": 0.82,
  "state": "APPROACHING",
  "trend": "increasing",
  "signal_pattern": "sustained_gradient",
  "source_proximity_hypothesis": "moderate",
  "recommended_action": "advance",
  "confidence": 0.78,
  "reasoning_summary": "Short-term signal ratio 1.34 and positive long-term slope (+0.021 ppm/frame) indicate movement toward source. Intermittency 0.55 consistent with structured plume contact. Confidence 0.78.",
  "window_size": 22,
  "features": {
    "short_trend": 1.34,
    "long_trend": 0.021,
    "intermittency": 0.55,
    "peak_snr": 4.1
  }
}

Confidence is derived from peak_snr × intermittency. Values below 0.5 indicate ambiguous state — treat recommended_action as tentative and pass this to the controller with reduced authority.

Examples

Signal loss — recovery casting

Window (VOC, normalized): [0.82, 0.91, 0.78, 0.12, 0.08, 0.05, 0.04, 0.06, 0.03, 0.04]
                                                  ↑ plume exit

Features: intermittency=0.30, peak_snr=1.1, short_trend=0.61
State → LOST_PLUME
recommended_action → "cast_upwind"
reasoning → "Intermittency dropped from 0.80 to 0.30 over 7 frames after prior contact.
             Begin crosswind casting with upwind bias. Expand search radius by 0.5 m
             per cast arm."

Near-source saturation

Window (VOC, normalized): [1.2, 1.8, 2.4, 2.9, 2.8, 3.1, 2.9, 3.0, 3.1, 2.8]

Features: intermittency=0.90, short_trend=1.02, peak_snr=1.4
State → NEAR_SOURCE
recommended_action → "slow_and_densify"
reasoning → "High sustained signal (mean 2.62 ppm normalized), intermittency 0.90,
             short-term ratio 1.02 indicates plateau. Likely within 1–3 m of source
             region. Reduce speed to 0.1 m/s; increase sampling to 10 Hz."

Turbulent contact in crosswind

Window (VOC, normalized): [0.0, 0.9, 0.0, 1.2, 0.0, 0.8, 0.0, 1.1, 0.0, 0.7]

Features: intermittency=0.50, peak_snr=5.2, short_trend=0.92
State → CONTACT_TURBULENT
recommended_action → "hold_and_cast"
reasoning → "Alternating high/zero pattern (peak_snr 5.2) indicates turbulent filament
             contact. Hold position; execute ±0.3 m lateral casts to locate filament
             center before advancing."

Constraints

Signal loss ≠ source absence — Turbulent plumes produce intermittent contact. LOST_PLUME triggers recovery casting, not mission abort.
Do not act on single spikes — One reading above threshold is noise until peak_snr > 2 and intermittency > 0.2 over at least 5 frames.
Absolute concentration is unreliable — Use baseline-corrected normalized values only. Raw ppm varies by temperature, humidity, sensor age, and cross-sensitivity.
State hysteresis is required — Three consecutive readings must confirm a new state before transitioning. Single-frame state flips indicate noise.
Sensor warm-up — Discard readings from the first 60 s after device power-on.
Wind data improves decisions — Without wind vector, casting is orthogonal to last heading. With wind vector, bias upwind using available ENV_* or anemometer data.
This skill produces recommendations, not commands — The downstream controller decides whether to execute. Always forward confidence so the controller can apply authority scaling (e.g., ignore actions when confidence < 0.4).
Plume Environment (simulation) — For training and evaluation, use the Scentience Plume Environment sandbox (gym.make("scentience/PlumeEnv-v0")), which provides ground-truth plume state for reward shaping.