olfactory-inertial-odometry/SKILL.md

metadata

name: olfactory-inertial-odometry
description: >
  Fuses Scentience OPU chemical signals with onboard IMU inertial measurements
  to produce drift-corrected position estimates in GPS-denied environments —
  olfactory inertial odometry (OIO). Use when navigating indoors, underground,
  or in RF-jammed environments; when localizing by persistent chemical
  landmarks; when evaluating OIO policies in the Scentience Navigation
  Environment; or when correcting IMU drift using scent as a positional anchor.
version: 1.0.0
author: kordelfrance
license: Apache 2.0
compatibility: >
  Hardware: Reconnaisscent (includes onboard IMU, BLE). UAV streaming via
  Scentience UAV Sockets API (wss://). Ground robots: standard Sockets API or
  ble-device skill. Simulation: Scentience Navigation Environment
  (gym-compatible, standard RL algorithms). Python: pip install scentience
  websockets gymnasium. Docs: scentience.github.io/docs-api.
metadata:
  domain: olfaction
  tags:
    - odometry
    - oio
    - navigation
    - imu
    - dead-reckoning
    - gps-denied
    - uav
    - robotics
    - sim2real
    - localization
  hardware:
    - Reconnaisscent
    - Scentinel
  depends_on:
    - ble-device
  apis:
    - UAV Sockets API
    - Sockets API
  related_papers:
    - >-
      Olfactory Inertial Odometry: Methodology for Effective Robot Navigation by
      Scent (IEEE Ubiquitous Robots)
    - >-
      Olfactory Inertial Odometry: Sensor Calibration and Drift Compensation
      (IEEE Inertial Sensors 2025)
    - >-
      Chasing Ghosts: A Sim2Real Olfactory Navigation Stack with Optional Vision
      Augmentation (arXiv 2026)
    - >-
      Chronoamperometry with Room-Temperature Ionic Liquids: Sub-Second
      Inference (IEEE BioSensors 2025)

Goal

Combine olfactory signal patterns with IMU dead-reckoning to localize a robot in environments where GPS is unavailable or unreliable. Persistent chemical features (concentration gradients, point-source plume trails) serve as position anchors that correct accumulated IMU drift — the olfactory equivalent of visual landmarks in VIO.

Instructions

1. Connect to the Data Stream

UAV / drone deployments — UAV Sockets API:

import asyncio, websockets, json

API_KEY = "SCN_..."
WS_URL  = "wss://api.scentience.ai/uav/stream"

async def stream_oio_frames():
    headers = {"Authorization": f"Bearer {API_KEY}"}
    async with websockets.connect(WS_URL, extra_headers=headers) as ws:
        async for message in ws:
            frame = json.loads(message)
            process_oio_frame(frame)

asyncio.run(stream_oio_frames())

Ground robots — standard Sockets API or BLE:

Use wss://api.scentience.ai/stream (same auth) or the ble-device skill. The BLE payload includes ENV_pressureHpa for barometric altitude; IMU data from the Reconnaisscent onboard IMU is delivered via the Sockets APIs.

2. OIO Frame Schema

Each fused frame from the UAV Sockets API includes OPU readings, IMU, and barometric altitude:

{
  "TIMESTAMP": 1743379200100,
  "UID": "SCN-REC-0042",
  "OPU": {
    "VOC": 0.45,
    "CO2": 412.0,
    "NH3": 0.08,
    "ENV_temperatureC": 22.4,
    "ENV_humidity": 48.1,
    "ENV_pressureHpa": 1013.2
  },
  "IMU": {
    "accel_x": 0.02, "accel_y": -0.01, "accel_z": 9.81,
    "gyro_x": 0.001, "gyro_y": -0.002, "gyro_z": 0.000,
    "mag_x": 24.1, "mag_y": -1.3, "mag_z": 43.2
  },
  "BARO": {
    "altitudeM": 12.4
  }
}

3. IMU Dead-Reckoning (Baseline Position Estimate)

Integrate linear acceleration to produce a raw position estimate. This drifts without olfactory correction:

import numpy as np

class DeadReckoning:
    """Simple double-integration dead-reckoning at fixed dt."""

    GRAVITY = np.array([0.0, 0.0, 9.81])

    def __init__(self, dt: float = 0.05):   # 20 Hz default
        self.dt       = dt
        self.position = np.zeros(3)
        self.velocity = np.zeros(3)

    def update(self, accel: np.ndarray) -> np.ndarray:
        linear = accel - self.GRAVITY          # remove gravity
        self.velocity += linear * self.dt
        self.position += self.velocity * self.dt
        return self.position.copy()

Note: heading error (gyro bias) accumulates separately. Use magnetometer fusion or visual odometry for heading correction alongside OIO position correction.

4. Define Olfactory Anchors (Chemical Landmarks)

An olfactory anchor is a persistent chemical feature at a known map coordinate. Examples: HVAC vent (NH3), industrial exhaust stack (SO2), refrigeration unit (NH3).

from dataclasses import dataclass
import numpy as np

@dataclass
class OlfactoryAnchor:
    compound: str           # e.g., "NH3"
    map_position: np.ndarray  # known 3D coordinate in map frame
    snr_threshold: float = 3.0  # minimum SNR to treat as valid detection
    noise_floor_ppm: float = 0.01

anchors = [
    OlfactoryAnchor("NH3", np.array([3.1, -1.0, 0.0]), snr_threshold=3.0),
    OlfactoryAnchor("NH3", np.array([18.4, 2.2, 0.0]), snr_threshold=3.0),
    OlfactoryAnchor("CO2", np.array([9.0, 0.0, 0.0]),  snr_threshold=4.0),
]

Anchors require at least 3 consecutive detections above threshold before triggering a correction. Single-frame detections are treated as noise.

5. Apply Olfactory Drift Correction

def apply_olfactory_correction(
    dr: DeadReckoning,
    frame: dict,
    anchors: list[OlfactoryAnchor],
    learning_rate: float = 0.3,   # partial correction weight (tune per environment)
    consecutive_required: int = 3,
    detection_counts: dict = None,
) -> tuple[np.ndarray, dict]:
    """
    Returns corrected position and updated detection_counts dict.
    learning_rate=0.3 means each anchor correction pulls 30% of the way to the anchor.
    """
    if detection_counts is None:
        detection_counts = {}

    opu = frame["OPU"]
    position = dr.position.copy()

    for anchor in anchors:
        raw_reading = opu.get(anchor.compound, 0.0)
        snr = raw_reading / anchor.noise_floor_ppm

        key = (anchor.compound, tuple(anchor.map_position))
        if snr >= anchor.snr_threshold:
            detection_counts[key] = detection_counts.get(key, 0) + 1
        else:
            detection_counts[key] = 0

        if detection_counts.get(key, 0) >= consecutive_required:
            # Confidence scales with SNR above threshold
            confidence = min(1.0, (snr - anchor.snr_threshold) / 10.0)
            error      = anchor.map_position - dr.position
            correction = learning_rate * confidence * error
            dr.position += correction
            position = dr.position.copy()

    return position, detection_counts

6. Run the Full OIO Loop

async def run_oio(anchors, dt=0.05):
    dr = DeadReckoning(dt=dt)
    detection_counts = {}

    async for frame in stream_oio_frames():
        imu   = frame["IMU"]
        accel = np.array([imu["accel_x"], imu["accel_y"], imu["accel_z"]])

        # Step 1: IMU dead-reckoning
        raw_position = dr.update(accel)

        # Step 2: Olfactory correction
        corrected_position, detection_counts = apply_olfactory_correction(
            dr, frame, anchors, detection_counts=detection_counts
        )

        # Step 3: Output
        output = {
            "timestamp_ms":       frame["TIMESTAMP"],
            "estimated_position": corrected_position.tolist(),
            "baro_altitudeM":     frame["BARO"]["altitudeM"],
        }
        emit(output)

7. Simulate and Train in the Navigation Environment

import gymnasium as gym

# Scentience Navigation Environment — gym-compatible, OIO scenario
env = gym.make("scentience/NavigationEnv-v0", scenario="oio_corridor")

obs, info = env.reset()
done = False
while not done:
    action = policy(obs)
    obs, reward, terminated, truncated, info = env.step(action)
    done = terminated or truncated

Supports PPO, SAC, TD3 and Sim2Real transfer. Use the oio_corridor scenario for single-compound anchors; oio_multi_source for multi-anchor environments.

8. Format OIO Output

{
  "timestamp_ms": 1743379215000,
  "estimated_position": {"x": 3.42, "y": -1.18, "z": 0.0},
  "position_confidence": 0.74,
  "drift_corrected": true,
  "imu_drift_estimate_m": 0.31,
  "last_anchor": {
    "compound": "NH3",
    "snr": 6.2,
    "map_position": {"x": 3.1, "y": -1.0, "z": 0.0},
    "correction_magnitude_m": 0.29
  },
  "recommendation": "maintain_heading",
  "reasoning": "NH3 anchor (SNR 6.2, 4 consecutive detections) applied 0.29 m correction. Position confidence 0.74. IMU drift prior to correction: 0.31 m."
}

Examples

GPS-denied indoor warehouse navigation

Environment: Indoor warehouse. NH3 from refrigeration units at 2 known positions.
Hardware: Reconnaisscent on ground robot. Polling at 20 Hz.
Anchors: 2 NH3 anchors at (3.1, -1.0) and (18.4, 2.2).

After 45 s without GPS:
  Raw IMU drift:               1.2 m
  Post-OIO position error:     0.18 m  (3 anchor correction events applied)
  Result: Robot navigated to target shelf within 0.2 m tolerance.

UAV outdoor plume tracking at altitude

Environment: Open field, wind 2.1 m/s NW, ethanol source at 40 m range.
Hardware: Reconnaisscent on UAV at 10 m altitude.
OIO anchor: Ethanol concentration gradient as a soft positional landmark.

Trajectory: UAV maintained <0.5 m altitude variance using barometric + OIO fusion.
            Reached within 1.8 m of source without GPS.
Sim2Real: Policy trained in NavigationEnv transferred directly to hardware with
          <12% performance degradation on source-reach success rate.

Constraints

• OIO corrects position drift, not heading — Gyro bias accumulates separately. Use magnetometer fusion or visual odometry for heading correction.
• Anchor persistence required — Require 3+ consecutive frames above SNR threshold before triggering a correction. Transient odor spikes do not qualify as landmarks.
• Prior map required for absolute localization — Without a map of anchor positions, OIO provides relative odometry only (drift reduction, not absolute localization).
• Turbulence degrades anchor reliability — Outdoor wind > 4 m/s disrupts plume structure. In high-wind conditions, increase snr_threshold and reduce learning_rate to avoid false corrections.
• Sensor calibration is prerequisite — Drift correction assumes per-compound baseline calibration. Run the calibration procedure from the IEEE Inertial Sensors 2025 paper before deployment. Uncalibrated sensors produce incorrect corrections.
• Sim2Real gap — The Scentience Navigation Environment uses simplified plume models. Real-world turbulence degrades policy performance; always validate on hardware after simulation training.
• Position estimates carry uncertainty — Implement fallback navigation (hover-and-wait, return-to-home) when position_confidence < 0.5. Never treat OIO output as a safety guarantee.
• Multi-compound disambiguation — When multiple anchor compounds are present, use the highest-SNR detection for correction. Do not average corrections across compounds with conflicting position hypotheses.