s2s-certify 1.5.0

Physics-certified motion data toolkit for Physical AI training

S2S — Physics-Certified Sensor Data

Physics-certified motion data for prosthetics, robotics, and Physical AI.

IMU sensor data is silently corrupted more often than people realize. S2S catches it using physics laws, not statistics. Proven on 5 real datasets. One line to install.

---

Live Demos

• 📊 Interactive Data Explorer — 104,160 real certified records, hover to explore
• 📱 Phone IMU Demo — real-time physics certification on your phone
• 🎥 Pose Camera Demo — 17-joint live certification

No install needed. All processing runs on your device. No data sent anywhere.

---

The Problem

Physical AI (robots, prosthetics, exoskeletons) is trained on motion data. But most datasets contain synthetic data that violates physics, corrupted recordings, and mislabeled actions — with no way to verify the data came from a real human moving in physically valid ways.

A robot trained on bad data learns bad motion. A prosthetic hand trained on uncertified data fails its user.

---

Four Proven Levels + Experimental AI

S2S improves model performance at every stage of the training pipeline. All results validated across five independent datasets at four different sampling rates.

Level 1 — Quality Floor ✅ PROVEN on 3 datasets

Dataset	Hz	Corruption	S2S Recovery	Net vs Clean
WISDM 2019	20Hz	35% corrupted	154% recovered	+1.74% F1
PAMAP2	100Hz	35% corrupted	confirmed	+0.95% F1
UCI HAR	50Hz	35% corrupted	135% recovered	+2.51% F1

Physics floor removes bad data and beats the clean baseline across three independent datasets at three different sampling rates.

Level 2 — Physics Quality Floor Generalises ✅ PROVEN on 3 datasets

Dataset	Hz	Data used	vs All data
WISDM 2019	20Hz	41% of windows	+1.74% F1
PAMAP2	100Hz	88% of windows	+0.95% F1
UCI HAR	50Hz	49% of windows	+2.51% F1

Also proven: kinematic chain consistency on PAMAP2 (hand + chest + ankle IMU):

Condition	F1	Δ
Single chest IMU	0.7969	baseline
3 IMUs naive concat	0.8308	+3.39%
3 IMUs + chain filter	0.8399	+0.91% over naive
Net vs single sensor	+4.23% F1	← headline

Less data, higher quality, better model. The physics score is a reliable proxy for training value — confirmed across devices, sampling rates, and activity types.

Level 3 — Biological Signal Certification ✅ PROVEN

Tested on PhysioNet PTT-PPG — 4 real subjects, 1164 windows, 500Hz wrist device, walk/sit/run.

Signal	Result
PPG pass rate	96.3% on real human subjects
Heart rate	mean 106 BPM (physiologically correct for activity)
HRV RMSSD	mean 21ms (real human variability)
Skin temperature	33.6°C (confirmed real human range)

Real pulse, real HRV, real temperature — verified simultaneously. Synthetic data cannot fake all three.

Level 4 — Multi-Sensor Fusion Coherence ✅ PROVEN

Dataset	HIL Score	Pass Rate	Tiers
PTT-PPG 500Hz wrist	68.7/100	100%	438 SILVER + 726 BRONZE
PAMAP2 100Hz (auto-Hz)	65.3/100	100%	87 SILVER + 13 BRONZE

Real sensors: PPG infrared + PPG red + IMU accel+gyro + skin temperature — all from the same wrist hardware.

If HR rose with activity, skin temperature stayed in human range, and IMU timing matched PPG — simultaneously — a human was there.

Level 5 — Physics-Informed Hybrid AI EXPERIMENTAL

Hybrid approach beats both baselines but physics feature extraction needs improvement.

Dataset	Model	Accuracy	Features	Status
PTT-PPG	Raw IMU	79.55%	768 features	Baseline
PTT-PPG	Physics Only	70.48%	19 features	Baseline
PTT-PPG	Hybrid	83.68%	787 features	+4.13% vs Raw, +13.20% vs Physics

Key Finding: Physics features add complementary signal but extraction is early-stage.

Feature Importance Analysis (Honest Assessment):

• Only 2/19 physics features contribute: rigid_rms_measured and resonance_peak_energy
• 17/19 physics features have zero importance in hybrid model
• Physics efficiency: 0.0032 per feature vs 0.0012 for raw (promising but needs work)
• Raw IMU dominates: 94% of predictive power from 768 features

What Works:

• rigid_rms_measured (0.0569 importance) - RMS acceleration magnitude
• resonance_peak_energy (0.0011 importance) - Frequency domain energy

What Needs Improvement:

• 17 physics laws produce non-predictive features for activity classification
• Most tier indicators (is_gold, is_silver, etc.) have zero importance
• Confidence scores and detailed law outputs not useful for ML

Status:

Physics feature extraction is early-stage but promising. The hybrid approach proves physics features add unique signal, but most physics laws need better feature engineering for ML tasks.

Next Phase: Improve physics feature extraction to make more laws ML-relevant.

---

Active Learning Pipeline ✅ PROVEN

Self-improving data quality system that learns from corruption patterns and generates training curriculum automatically.

Module 1 — Corruption Fingerprinter

• Purpose: Detect and classify data corruption types
• Results: Identified resonance_frequency as most vulnerable (77% of corruptions break it first)
• Status: ✅ Proven on PTT-PPG data

Module 2 — Frankenstein Mixer

• Purpose: Find exact contamination boundaries for each physics law
• Results: IMU consistency breaks at 30.6% contamination, resonance at 29.2%, jerk at 53.7%
• Status: ✅ Proven on PTT-PPG data

Module 3 — Curriculum Generator

• Purpose: Generate training data at every quality level automatically
• Results: 2,000 samples with balanced tiers (GOLD 5.7%, SILVER 58.7%, BRONZE 28.1%, REJECTED 6.7%)
• Auto-discovery: Found NinaPro DB5, EMG Amputee, HuGaDB, PTT-PPG automatically
• Status: ✅ Proven on Mac with auto-discovery

Module 4 — Cloud Trainer

• Purpose: Train quality prediction models on curriculum data
• Results: 85.5% accuracy (+27.7% over 57.8% majority baseline)
• Best model: GradientBoosting with 93% precision on SILVER tier
• Status: ✅ Proven on Mac with sklearn baseline

Pipeline Impact:

• Automatic curriculum generation from any local dataset
• Quality predictor that significantly outperforms naive baseline
• Self-improving system that learns corruption patterns
• Ready for deployment with trained models and prediction API

---

Auto-Hz Device Detection

S2S automatically detects device profile from two numbers already in the data — sampling Hz (from median timestamp intervals) and signal amplitude range (from first window). No user configuration needed.

Hz range	Signal range	Profile	Example
≥400Hz	<1.0 normalized	normalized_500hz	PTT-PPG
≤150Hz	>10 raw ADC	raw_adc_100hz	PAMAP2
other	other	default	fallback

Before auto-Hz: PAMAP2 Level 4 HIL = 38.4. After: 65.3. Same data, correct profile.

---

Validated on Real Human Data

WISDM 2019 (51 subjects, 20Hz, wrist accel, 18 activities):

Level	Result
Level 1	+1.74% F1 vs corrupted, 154% recovery
Level 2	+1.74% F1 vs all data, 41% of windows used

PAMAP2 (9 subjects, 100Hz, hand+chest+ankle IMU, 12 activities):

Level	Result
Level 1	+0.95% F1 vs corrupted
Level 2	+4.23% F1 kinematic chain vs single sensor
Level 4	HIL 65.3/100, 100% pass, 87 SILVER

UCI HAR (30 subjects, 50Hz, body accel+gyro, 6 activities):

Level	Result
Level 1	+2.51% F1 vs corrupted, 135% recovery
Level 2	+2.51% F1 vs all data, 49% of windows used

PhysioNet PTT-PPG (4 subjects, 500Hz, wrist PPG+IMU+thermal, walk/sit/run):

Level	Result
Level 2 IMU	61.7% pass rate, avg score 37.2/100
Level 3 PPG	96.3% pass rate, HR 106 BPM, HRV 21ms
Level 4 Fusion	HIL 68.7/100, 100% pass, 438 SILVER

NinaPro DB5 (10 subjects, 2000Hz, 16-channel EMG + 3-axis accelerometer, hand gestures):

Level	Result
Law 1 Newton	EMG→accel lag 117.5ms mean, 81.6% in 50–200ms range, 10/10 subjects

Muscle fires → limb accelerates 117.5ms later. Consistent with published neuromuscular literature. Shuffled baseline: 88.5ms — real causal lag is distinct. Synthetic data cannot reproduce without full rigid-body muscle simulation.

---

11 Physics Laws

Single-Sensor Laws (Levels 1–3)

#	Law	What It Catches
1	Newton's Second Law (F=ma, 117.5ms EMG→accel lag)	Synthetic data missing lagged EMG-accel correlation
2	Segment Resonance (ω=√(K/I))	Tremor at impossible frequency for body segment
3	Rigid Body Kinematics (a=α×r+ω²×r)	Gyro and accel generated independently
4	Ballistocardiography (F=ρQv)	IMU missing cardiac recoil
5	Joule Heating (Q=0.75×P×t)	Sustained EMG without thermal elevation
6	Motor Control Jerk (∂³x/∂t³ ≤ 5000 m/s³)	Robotic or keyframe animation artefacts
7	IMU Consistency (Var(accel) ~ f(Var(gyro)))	Accel and gyro from independent generators

Multi-Sensor Chain Laws (Level 4)

#	Law	What It Catches
8	Locomotion Coherence (freq spread <2.5Hz)	Sensors recording different activities
9	Segment Coupling (chest-ankle r >0.3)	Independent synthetic channels
10	Gyro-Accel Coupling (per IMU)	Rotation without corresponding acceleration
11	Cross-Sensor Jerk Timing (ankle leads chest 0–200ms)	Reversed or zero lag — not real heel-strike

---

Tier System

Tier	Score	Meaning
GOLD	≥87	All physics laws passed. Pristine.
SILVER	75–86	Trusted. Minor deviations within noise.
BRONZE	60–74	Marginal. Candidate for reconstruction at ≤50Hz.
RECONSTRUCTED	—	Repaired, re-scored ≥75, spectral sim ≥0.8. Weight 0.5.
REJECTED	<floor	Removed from pipeline.

Floor = p25 of clean score distribution per dataset (adaptive).

---

Live API

No install needed:

curl -X POST https://s2s-65sy.onrender.com/certify -H "Content-Type: application/json" -d "{"accel": [[ax,ay,az],...], "sample_rate_hz": 50}"

import requests
cert = requests.post("https://s2s-65sy.onrender.com/certify", json={"accel": data, "sample_rate_hz": 50})
print(cert.json()["tier"])  # GOLD / SILVER / BRONZE / REJECTED

Or install locally:

Install

pip install s2s-certify

Zero dependencies. Pure Python 3.9+. Works on any platform.

---

Quick Start

from s2s_certify import certify

result = certify(accel_window, sample_rate_hz=20)

print(result['tier'])        # GOLD / SILVER / BRONZE / REJECTED
print(result['score'])       # 0–100
print(result['laws_passed']) # which physics laws passed

s2s-certify your_imu_data.csv
s2s-certify your_imu_data.csv --output report.json

---

Datasets Validated

Dataset	Hz	Sensors	Windows	Used for
WISDM 2019	20Hz	Wrist accel	46,946	Levels 1, 2
PAMAP2	100Hz	Hand+Chest+Ankle IMU	13,094	Levels 1, 2, 4
UCI HAR	50Hz	Body accel+gyro	10,299	Levels 1, 2
PhysioNet PTT-PPG	500Hz	Wrist PPG+IMU+Thermal	1,164	Levels 2, 3, 4, 5 (experimental)
NinaPro DB5	2000Hz	Forearm EMG+Accelerometer	500	Law 1

---

Paper

S2S: Physics-Certified Sensor Data — Four Proven Levels, Eleven Laws, Five Independent Datasets

→ Read paper (PDF) | → DOI: 10.5281/zenodo.18878307

---

Project Structure

s2s_standard_v1_3/     # Physics engine (zero dependencies)
experiments/           # All experiments + results JSON
tests/                 # 110 tests, all passing
docs/paper/            # S2S_Paper_v5.pdf

---

License

BSL-1.1 — free for research and non-commercial use. Contact for commercial licensing.