OmniRay Avx2 : A CPU-Based Autonomous Active SLAM Framework with Adaptive Curriculum Learning and AVX2 Acceleration

A pluggable SIMD raycasting engine, vectorized particle filter, and Gymnasium environment for training and benchmarking Active SLAM agents on consumer CPU hardware.

Author: Kingshuk Chatterjee  |  Co-Author: Ayush Ranjan  |  Co-Author: Raghav Singh Parihar

---

Overview & Architecture

OmniRay integrates a 256-bit C++ AVX2 SIMD raycasting backend, a vectorized NumPy particle filter (VectorSLAM), and a Gymnasium environment with sim-to-real noise models. It enables end-to-end Deep Reinforcement Learning (DRL) for Active SLAM on consumer CPUs without dedicated GPUs.

---

Hardware Disclosure

All reported performance metrics were measured on a consumer ultrabook:

Component	Specification
Device	ASUS Zenbook S13 OLED
CPU	Intel Core i7-1355U (10-core / 12-thread, hybrid P+E architecture)
GPU	None (integrated Intel Iris Xe Graphics only)
RAM	16 GB LPDDR5x, 5200 MHz
SIMD ISA	AVX2 (256-bit, 8-lane parallel raycasting)
OS / Toolchain	Python 3.11; C++ backend compiled via CMake (Release)

---

5-Layer Self-Adaptive Autonomy System

OmniRay features a 5-layer hierarchical feedback control architecture for real-time autonomy adaptation:

Layer	Module	Function
1	health_monitor.py	System State Monitoring — Computes health score $H_t \in [0, 1]$ from entropy rate, coverage velocity, and SLAM divergence.
2	adaptive_reward.py	Dynamic Reward Adaptation — Adjusts reward coefficients in real time based on system health.
3	meta_policy.py	Meta-Policy Controller — Neural controller optimizing reward parameters via policy gradient updates.
4	curriculum.py	Curriculum Manager — Dynamically adjusts map difficulty, noise, and step budgets.
5	continual_learner.py	Continual Learner — In-deployment replay buffer with online updates and checkpoint/rollback safety.

---

Quick Start

1. Install Dependencies & Build SIMD Core

pip install -r requirements.txt

# Compile C++ SIMD backend
cd sim
mkdir build
cd build
cmake ..
cmake --build . --config Release
cd ../..

2. Run Agent Visualizer & Profiler

# Visualize pretrained agent
py -3.11 visualize_agent.py --model-path active_slam_ppo_robust_master.zip --episodes 3

# Benchmark backends & test environment
py -3.11 -m profiling.benchmark_bottleneck --rays 360 --iterations 500
py -3.11 test_env.py --backend simd --episodes 3

3. Adaptive Training

py -3.11 train_rl.py --adaptive --meta-policy --curriculum --continual --total-steps 100000

---

Quantitative Benchmarks

1. Raycasting Core Latency (360 Rays)

Backend	Mean Scan Time	Median Scan Time	P99 Scan Time	Estimated Time (100K Steps)	Performance Profile
Pure Python	6.139 ms	6.469 ms	7.636 ms	10.2 min	Baseline
PyMunk (segment_query)	3.009 ms	3.447 ms	4.785 ms	5.0 min	Moderate
NumPy (vectorized)	0.225 ms	0.262 ms	0.373 ms	0.4 min (24 s)	Fast
C++ SIMD (AVX2)	0.038 ms	0.038 ms	0.089 ms	0.06 min (3.8 s)	Ultra-Low Latency (~26× speedup)

2. Fixed Evaluation Suite & Multi-Sequence Benchmark

Sequence / Environment	ATE RMSE (m)	Loop Closure Failures	Peak RAM (MB)	Scalar CPU Latency ($\mu\text{s}$)	AVX2 SIMD Latency ($\mu\text{s}$)	Hardware Speedup
Intel Research Lab (Seattle)	0.042	0	41.8 MB	962 $\mu\text{s}$	37.1 $\mu\text{s}$	25.9×
MIT Stata Center	0.058	0	54.2 MB	1,420 $\mu\text{s}$	52.4 $\mu\text{s}$	27.1×
Freiburg Building 52	0.039	0	38.5 MB	810 $\mu\text{s}$	31.8 $\mu\text{s}$	25.5×
Holdout Set: ACES (UT Austin)	0.064	1	62.1 MB	1,680 $\mu\text{s}$	61.2 $\mu\text{s}$	27.5×

3. Multi-Model Baseline Comparisons

Intel Research Lab Floorplan:

Model / Algorithm	Coverage (%)	Path Length (m)	Coverage / Meter (%/m)	Decision Latency (ms)	Peak RAM (MB)	Collision Count
Random Walk	65.55 ± 12.52%	148.57 ± 67.34 m	0.4411 %/m	7.16 ms	34.2 MB	225.7 ± 33.7
Yamauchi (1997)	70.85 ± 14.67%	155.06 ± 64.90 m	0.4570 %/m	8.61 ms	39.5 MB	192.3 ± 39.9
RRT-Exploration (2017)	85.55 ± 13.57%	267.64 ± 33.16 m	0.3200 %/m	10.49 ms	44.8 MB	30.7 ± 28.2
Stachniss (2005)	95.43 ± 0.07%	207.65 ± 22.06 m	0.4600 %/m	18.51 ms	52.1 MB	0.0 ± 0.0
OmniRay (Ours)	90.87 ± 3.29%	418.12 ± 106.25 m	0.2172 %/m	3.09 ms	41.8 MB	62.3 ± 46.5

MIT Stata Center Zero-Shot Benchmark:

Model / Algorithm	Coverage (%)	Path Length (m)	Coverage / Meter (%/m)	Decision Latency (ms)	Peak RAM (MB)	Collision Count
Random Walk	59.81 ± 19.40%	140.14 ± 96.38 m	0.4268 %/m	7.10 ms	35.1 MB	230.0 ± 48.3
Yamauchi (1997)	91.04 ± 4.40%	281.62 ± 37.18 m	0.3233 %/m	10.29 ms	41.2 MB	28.3 ± 20.5
RRT-Exploration (2017)	95.08 ± 0.07%	260.61 ± 2.99 m	0.3649 %/m	10.77 ms	46.5 MB	35.0 ± 4.5
Stachniss (2005)	95.33 ± 0.36%	205.22 ± 8.51 m	0.4654 %/m	11.96 ms	53.8 MB	1.7 ± 2.4
OmniRay (Ours - Zero Shot)	91.08 ± 3.99%	454.34 ± 108.03 m	0.2005 %/m	3.09 ms	41.8 MB	70.3 ± 53.6

Zero-Shot Generalization Gap:

Dataset / Environment	Environment Type	Final Coverage (%)	Zero-Shot Generalization Gap
Intel Research Lab (Seattle)	Fixed Training / Benchmark Floorplan	90.87 ± 3.29%	Baseline reference
MIT Stata Center (MIT dataset.bag)	Unseen Complex Atrium Holdout	91.08 ± 3.99%	+0.21%

---

Sim-to-Real Noise Robustness & Ablation Studies

Reward Specification

Parameter	Default	Description
reward_exploration	$1.0$	Reward per newly explored grid cell.
reward_time_penalty	$0.01$	Per-step efficiency penalty.
reward_collision_penalty	$0.1$	Collision penalty.
reward_frontier	$0.1$	Vectorized reward for frontier heading alignment.

Robust Evaluation & Exploration Progression

Multi-Seed Ablation Matrix (14 Configurations, 3 Seeds)

---

Repository Structure

OmniRay/
├── ablation_eval_full/  # Output plots and 3-seed ablation matrix charts
├── assets/              # Architecture diagrams
├── envs/                # Gymnasium environment, VectorSLAM, & 5-Layer Autonomy
├── profiling/           # Raycasting and SLAM profiler scripts
├── sim/                 # C++ 256-bit AVX2 SIMD raycasting engine & pybind11 bindings
├── config.yaml          # Hyperparameters and network configs
├── train_rl.py          # PPO training pipeline
├── visualize_agent.py   # Environment visualization tool
└── evaluate_and_record.py # Quantitative evaluation runner

---

Authors & License

• Author: Kingshuk Chatterjee
• Co-Author: Ayush Ranjan
• Co-Author: Raghav Singh Parihar
• Hardware Dependency: Latencies reflect AVX2 execution on Intel i7-1355U. Unsupported hardware falls back to NumPy mode.
• License: Apache License 2.0. Copyright 2026 Kingshuk Chatterjee, Ayush Ranjan, Raghav Singh Parihar.