diff --git a/intel/presentations/assets/pipeline_bg.png b/intel/presentations/assets/pipeline_bg.png new file mode 100644 index 0000000..2403497 Binary files /dev/null and b/intel/presentations/assets/pipeline_bg.png differ diff --git a/intel/presentations/assets/title_bg.png b/intel/presentations/assets/title_bg.png new file mode 100644 index 0000000..8c2c0cc Binary files /dev/null and b/intel/presentations/assets/title_bg.png differ diff --git a/intel/presentations/presentation.html b/intel/presentations/presentation.html new file mode 100644 index 0000000..5dfb044 --- /dev/null +++ b/intel/presentations/presentation.html @@ -0,0 +1,342 @@ + + + + + + Fox CARLA ADAS Simulation Platform + + + + + + + +
+ +
+

High-Fidelity Deterministic
Radar Simulation

+

Moving from bounding-box emulation to mathematically pure, hardware-configurable physics validating the Shenron Pipeline.

+
+
+ + +
+

The Physical Foundation: 1/R⁴ Physics

+
+
+

Standard simulator bounding-boxes bypass the complex realities of electromagnetic propagation. The Shenron engine enforces strict free-space physics for absolute accuracy.

+ +
+

Milestone R4: Pure Physics

+

We abandoned artificial density and distance normalizations. All generated scatterers now strictly follow the Radar Range Equation (1/R⁴). A building at distance naturally drops in power compared to a vehicle up close.

+
+ +
+

Deterministic Coordinate Control

+

Standard Traffic Manager leads to high variance. Using a Spawn-and-Move logic with enforced Velocity overrides guarantees complex scenarios (like left-turn cut-ins) occur at the exact same millisecond every run.

+
+
+
+
+ +
Insert Screenshot: 1/R⁴ Contrast
+
Provide a side-by-side or CFAR output showing how distant heavy clutter is naturally attenuated while the closer vehicle is correctly highlighted due to the inverse square law path loss.
+
+
+
+
+ + +
+

Unprecedented Configurability: "Knobs & Dials"

+
+
+

The Shenron architecture eliminates hardcoded physical scalars. Instead, it exposes a "Control Panel" mimicking real hardware tuning. If we do it on the test track, we can do it in the simulation.

+ +

Supported Hardware Profiles

+
+ TI AWRL1432 (ADAS) + Radarbook (Research) + TI Cascade (MIMO) +
+ +
    +
  • Bandwidth (B) & Chirps: Accurately adjust range and doppler resolution directly reflecting radar specifications.
    • +
  • System Gain Calibration (110dB): Modify the global signal chain gain ($P_t G_t G_r$) without touching the core physics engine.
    • +
  • Antenna Arrays (nRx/nTx): Instantly swap between 6 virtual receivers (1432) or 8 virtual receivers (Radarbook) to measure DSP beamforming limits.
    • +
  • Noise Amplitude: Dynamically inject AWGN thermal noise to test CFAR floor robustness.
    • +
+
+
+
+ +
Insert Code/UI Screenshot: Customizing Radar
+
Please insert a screenshot showing `ConfigureRadar.py` or the specific JSON payload being passed via the dashboard to demonstrate how easily we swap between a 1432 profile and a Radarbook profile.
+
+
+
+
+ + +
+

Deep Observability: The Metrology Suite

+
+
+

Instead of treating synthetic radar output as a "black box," our new pipeline visualizes the entire signal chain. This allows engineers to cross-validate ADC output against physical ground truth.

+ +
+

The "Radar Blue" Standard

+

We instituted phase-preserving Doppler-Slice Synthesis to replace artifact-heavy coherent integration, resulting in pristine, sharp heatmaps.

+
+ +
    +
  • Range-Doppler (RD): Extracts target separation and velocity, preventing velocity zeroing.
    • +
  • Range-Azimuth (RA): Confirms azimuthal accuracy utilizing Zero-Tilt Symmetry (120° FOV perfectly aligned to boresight).
    • +
  • CFAR Gate: Visualizes adaptive thresholds distinguishing targets from the synthesized noise floor.
    • +
+
+
+
+ +
Insert Screenshot: 3-Plot Metrology Suite
+
Insert a screenshot of the Foxglove layout showing the 3 core metrology graphs (RD, RA, and CFAR) running alongside the 3D Pointcloud view to illustrate deep observability.
+
+
+
+
+ + +
+

Major Breakthroughs & Bug Squashing Engine

+

Achieving this level of physical simulation required identifying and tearing down several legacy miscalculations that corrupted early models.

+ +
+
+

The Isotropic Illumination Fix

+

The deepest physics problem resolved. Early simulation allowed side-targets at 80° to return 100% power, masking forward vehicles.

We implemented Horizontal (Azimuth) & Vertical (Elevation) Directional Gain Patterns. Clutter at wide angles is now correctly attenuated by ~15-20 dB, letting the true target dominate.

+
+
+
+

The Cos(Cos) Reflection Error

+

Recovered 46% of lost signal energy by tracking down a double-cosine application in the Open3D coordinate translation pipeline.

+
+
+
+

Floating Point Data Packing

+

Resolved the infamous "0.0 velocity" bug by utilizing `np.view(np.uint32)` to correctly deserialize perfectly packed 32-bit semantic tags.

+
+
+
+
+ + +
+

The Full Automation Pipeline

+

A completely automated workflow enabling one-click testbenches, fully orchestrated via a web dashboard.

+ +
+ +
Insert Graphic/Screenshot: Pipeline & Dashboard
+
+ Place a diagram or screenshot here showing the flow:

+ 1. GUI Dashboard (with Hardware GPU Idle Mode) → 2. CARLA Simulator → 3. Shenron Physics → 4. MCAP Serializer → 5. Foxglove Studio +
+
+ +
+

From deterministic scenario configuration to a 3D Foxglove validation pointcloud — without manual intervention.

+
+
+ + + diff --git a/intel/presentations/presentation_3_slide_navy.html b/intel/presentations/presentation_3_slide_navy.html new file mode 100644 index 0000000..70adc26 --- /dev/null +++ b/intel/presentations/presentation_3_slide_navy.html @@ -0,0 +1,446 @@ + + + + + + 3 Slide Fox ADAS Presentation + + + + + + + +
+

Deterministic ADAS Simulation

+ +
+ + +
+
+ [ Insert Image: CARLA RGB/Game View ] +
CARLA Simulation
+
+
+ [ Insert Image: Radar Heatmap (Foxglove) ] +
Shenron Radar
+
+
+
+
+ + + +
+

Shenron: Physics-Based Radar Engine

+ +
+
+
CARLA & Synchronizer
+
[ Graphic or Logo ]
+
+ +
+ +
+
Shenron Physics Engine
+
[ CPU/GPU Graphic ]
+
+ +
+ +
+
Radar Signal Processing
+
[ Heatmap / DSP ]
+
+ +
+ +
+
ADAS-Ready Output
+
[ Point Cloud / MCAP ]
+
+
+ +
+
+
1/R⁴ Accurate Spatial Decay Physics
+
+
Dropped artificial density variables. Reflected energy now strictly attenuates by the inverse square law, forcing massive background structures to cede dominance to near-field targets.
+ +
+
Directional Antenna Pattern Modeling
+
+
Resolved the "Isotropic Illumination" limit by implementing true Azimuth & Elevation gain logic. Radar blind-spots act realistically, attenuating 80° side-wall clutter by ~20dB.
+ +
+
Modular Hardware-Matched Pipeline
+
+
The "Knobs and Dials" philosophy completely separates base physics from simulation parameters. Easily swap profiles between AWRL1432, Radarbook, and MIMO Cascade hardware.
+
+
+ + + +
+
+

From Simulation to Validation

+
+ +
+ +
+
+

Physics-Driven
Metrology Accuracy

+
    +
  • Zero-Tilt Symmetry guarantees 120° FOV precision right down the boresight.
    • +
  • Doppler-Slice phase continuity resolves standard blur, delivering sharp "Radar Blue" PointCloud clusters.
    • +
+
+ + +
+
+

Real-World Hardware
Calibration Logic

+
    +
  • 110dB System Gain tuning aligns exact radar spec hardware profiles.
    • +
  • Modifiable Bandwidth (B), Chirp Repetition, and virtual receiver sets to replicate DSP behavior dynamically.
    • +
+
+ + +
+
+

Automated Core
Testing Platform

+
    +
  • GUI Dashboard orchestrating idle-mode GPU utilization for Shenron rendering safely.
    • +
  • Direct pipeline translating serialized JSONL/NPY data into FOXGLOVE MCAP standards.
    • +
+
+
+ +
+ Bridging the Gap from Research to Real-World Deployment +
+
+ + + diff --git a/scripts/ISOLATE/e2e_agent_sem_lidar2shenron_package/shenron/Sceneset.py b/scripts/ISOLATE/e2e_agent_sem_lidar2shenron_package/shenron/Sceneset.py index 2fb17ef..628d89f 100644 --- a/scripts/ISOLATE/e2e_agent_sem_lidar2shenron_package/shenron/Sceneset.py +++ b/scripts/ISOLATE/e2e_agent_sem_lidar2shenron_package/shenron/Sceneset.py @@ -240,7 +240,20 @@ class Sceneset(): # o3d.visualization.draw_geometries([pc]) loss_att,_,_ = get_loss_3(self.rad_scene, rho, az_boresight, elev_angle, angles_carla, radar, use_spec=False, use_diffused=True, no_material=False) - return rho, theta, loss_att, speed, angles + # --- Iteration 38: Phase Randomization (Millimeter Jitter) --- + # Apply physical roughness to distance to break impossible coherent summation + # (The Perfect Mirror effect) on mathematically flat CARLA meshes. + material_idx = np.asarray(self.rad_scene[:, 4], dtype='int') + + # Roughness mapping in meters (Aligned with get_loss_3 physical traits) + # 0=Unlw, 1=Wood(5mm), 2=Conc(5mm), 3=Hum(10mm), 4=Metal(50um) + roughness_map = np.array([0.0, 0.005, 0.005, 0.01, 0.00005]) + jitter_std = roughness_map[material_idx] + + # Randomly shift the distance for each point by its physical roughness magnitude + rho_jittered = rho + np.random.normal(loc=0.0, scale=jitter_std) + + return rho_jittered, theta, loss_att, speed, angles def get_loss(points, rho, angles, radar, use_spec = True, use_diffused = True, no_material = False): @@ -447,9 +460,13 @@ def get_loss_3(points, rho, az_boresight, elev_angle, angles, radar, use_spec = phi_deg = np.rad2deg(np.abs(np.pi / 2 - elev_angle)) G_ant = np.exp(-2.77 * np.power(phi_deg / radar.vertical_beamwidth, 2)) + # --- Iteration 37: Area Integration (Resolution Independence) --- + # A single LiDAR point represents an expanding physical patch of Area = R^2 * dTheta * dPhi + point_area = np.power(rho, 2) * voxel_theta * voxel_phi + # --- Iteration 17 preserved: Pure Physical 1/R^2 Tx path loss --- - # Each LiDAR point acts as a raw unit scatterer. No legacy density normalization. - P_incident = (1 / np.power(rho, tx_dist_loss_exponent)) * K_sq * G_ant + # Intercepted power is weighted by the physical area the point represents + P_incident = (1 / np.power(rho, tx_dist_loss_exponent)) * K_sq * G_ant * point_area # DEBUG: Monitor Signal Trends # P_inc print suppressed — data captured via model.get_signal_metrics() telemetry