Sensorflow Studio

Dataset Configuration

Prepare and validate your image source path and annotation files.

Strict Execution

Dataset Configuration Settings

Dataset Type / Source

Data Configuration YAML Contains directories for train, val, and class labels.

Images Path (Inference Source) Folder containing raw images for inference/auto-labeling.

What actually happened?

No execution yet — Load & Preprocess to generate verifiable evidence.

Workspace Status

Base Workspace: /DrivingRepo

Train script: Verifying…

Infer script: Verifying…

Auto-Grader script: Verifying…

Backend health: —

Dataset Configuration Editor

Edit active dataset configurations (classes, paths) directly in the YAML file.

📊 Dataset Row Statistics & Technical Metadata

Catalog reference stats for the selected dataset type. This is not proof that frames are on disk or browsable — use Validate & Browse or Pipeline Outputs.

Catalog estimate — not browsable

Source Rows in Dataset

—

Catalog estimate (not disk)

Rows Ingested / Loaded

—

Catalog estimate (not disk)

Ingestion Sampling Rate

—

Catalog estimate only

Total Labeled Objects

—

Catalog estimate only

Sensor & Capture Metadata

Sensor Modality:Monocular Front RGB Camera

Geographic Location:Local Drive

Weather & Lighting:Daytime / Mixed

Annotation Methodology:Auto-Labeler + Manual

Storage & Format Specs

Storage Footprint:6.5 MB

File Format:YOLOv8 TXT / COCO YAML

Licensing Terms:Ultralytics AGPL-3.0

Dataset Type ID:local

Class Object Frequencies (Loaded Rows)

Exabyte Ingest, Fusion & Stratification

Ingest one vendor by default (homogeneous frame IDs). Or use Load all datasets to register Local + each AV vendor as separate sequences under runs/pipeline/<prefix>_<vendor>/. Missing lakes become clearly labeled demo stubs (or NOT_EXECUTED if stubs are disabled).

Vendors

Local frames / video Alpamayo (demo stub) Waymo (demo stub) Audi A2D2 (demo stub)

Mix vendors (fuse multiple stubs into one sequence) Allow demo stubs when AV lakes are missing

Local source path (images folder or video file) Uses Dataset Configuration Images Path when left as default. Enable Local above to load all frames (vendor stubs alone are ~3 frames each).

Waymo root (optional)

Alpamayo root (optional)

A2D2 root (optional)

Max frames (local / lake folders)

Sequence ID / prefix Single ingest uses this ID. Load all datasets writes <prefix>_local, <prefix>_waymo, etc.

Pipeline Outputs — browse & validate

Inspect ingested frames, proposal JSON, and run artifacts under runs/pipeline/<sequence>/. Catalog “100% loaded” does not mean outputs are browsable here.

Sequence

Select a sequence or run ingest to browse outputs.

FRAMES

No frame selected.

Automated 3D Perception (SAM + LiDAR)

Generate 2D SAM masks and lift to 3D bounding box proposals.

SAM Checkpoint

Default path models/sam_vit_b.pth. On Hugging Face Space, upload weights or run locally with GPU.

Device

Skip SAM (fallback / NOT a model SUCCESS)

What actually happened?

No auto-label execution yet.

Pipeline Outputs — proposals & frames

After auto-label, open a frame to validate proposal JSON written under runs/pipeline/<sequence>/proposals/.

Sequence

Run auto-label, then refresh to browse proposals.

FRAMES

No frame selected.

Temporal Tracking & Consistency

Kalman + Hungarian association for ID-smooth multi-frame tracks.

What actually happened?

Ledger wiring pending — status will show NOT_EXECUTED until this stage records evidence.

Quality Gate & Comparative Analytics

Benchmark against vendor GT: mAP, orientation error, ID swaps, process_units.

What actually happened?

Evidence panel stub — gate results are live, full ledger wiring deferred.

Launch Gate Validation

Verify safety thresholds before export is allowed.

Base Model Selection

⚡

YOLOv8 Nano (n)

Fastest, ultra-lightweight. Ideal for local testing and mobile/edge devices.

6.5 MB

🚀

YOLOv8 Small (s)

Good trade-off between throughput and accuracy.

22.5 MB

🧠

YOLOv8 Medium (m)

Recommended standard. Balanced accuracy/speed profile.

52.1 MB

🌌

Alpamayo-R1-10B

Reasoning-focused autonomous driving model with Chain-of-Causation traces.

10B Params

🏎️

Alpamayo-1.5-10B

High-performance trajectory generation and lane-following control model.

10B Params

Execution Device

Compute Device

NVIDIA DGX Spark Remote Cluster Setup:

Follow these commands to configure the libraries on your remote node:

# 1. Start Tailscale on the DGX node & login
sudo tailscale up

# 2. Setup python virtual environment and activate
python3 -m venv .venv
source .venv/bin/activate

# 3. Install PyTorch with CUDA 12 support
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121

# 4. Install Ultralytics and Hugging Face dependencies
pip install ultralytics huggingface_hub

# 5. Enable Ollama server host mapping for remote Gemma 4:
export OLLAMA_HOST=0.0.0.0
ollama run gemma4:26b

Selected weights: yolov8n.pt

Training Controls

Epochs

Batch Size

What actually happened?

No training execution yet.

Real-time Output Logs & Curves

Waiting for training to start...

Auto-Labeler Parameters

Weights Path

Confidence Threshold 0.25

IoU NMS Threshold 0.45

What actually happened?

No inference execution yet.

Interactive Predictions Explorer

Image Select:

Quality Score Summary

-- Quality Score

-- Images

-- Detections

What actually happened?

No grader execution yet.

Issue Diagnostics & Recommendations

No issues scanned. Run auto-grader diagnostics.

🤖

AI Triage Copilot

Deep audit of the pipeline, data, and quality issues using Gemma 4 (26B)

No analysis generated yet. Click "Ask AI Triage Copilot" to run the LLM evaluation.

Multi-Camera Viewport

FRONT_CAM

NuRec Telemetry & Kinematics

GPS Position: --

Velocity: --

Acceleration: --

IMU Orientation (P/R): --

Human Bounding Box Triage

ID	Class Label	Box Coordinates [x, y, w, h]	Conf	Action

Chain-of-Causation (CoC) Reasoner

Reasoning / Decision Explanation Path

Primary Model (A)

Compare Model (B)

Side-by-Side Annotation Boundary Comparators

YOLOv8-M

Alpamayo-R1

Visual annotation overlay comparing bounding box predictions at frame coordinate intersections.

Key Safety & Quality Metrics

SOTA Annotation Metrics Reference

Annotation Model	Type	Latency (ms)	mAP [50]	Risk-Weighted Recall	Recall @ Critical Distance	VRU Recall (Pedestrians)	CoC Traces

Export Deployment Engine

Serialize your trained model weights for production deployment and inference acceleration.

Target Format

ONNX Engine (Standard CPU/GPU) TensorRT (.engine)

Finetuned Weights Details

Best Weights File: runs/detect/coco_finetuned/weights/best.pt

Last Weights File: runs/detect/coco_finetuned/weights/last.pt

IDE Model Context Protocol Servers

Enable or disable active plugins and external compute layers directly in mcp_config.json.

Raw config file: `mcp_config.json`

Path: ~/.gemini/config/mcp_config.json

0 Total Conflicts

0 Critical

0 High

0 Medium

0 Low

California Statewide Conflict Map

Critical High Medium Low

Conflict Analysis Data Grid (FHWA SSAM)

Street Interface ⇅	County ⇅	Type	Min TTC ⇅	Min PET ⇅	Max ΔS ⇅	Severity ⇅

Page 1 / 1

Dataset Configuration Settings

Official Dataset Access Portals:

What actually happened?

Workspace Status

Dataset Configuration Editor

📊 Dataset Row Statistics & Technical Metadata

Sensor & Capture Metadata

Storage & Format Specs

Class Object Frequencies (Loaded Rows)

Exabyte Ingest, Fusion & Stratification

Pipeline Outputs — browse & validate

Automated 3D Perception (SAM + LiDAR)

What actually happened?

Pipeline Outputs — proposals & frames

Temporal Tracking & Consistency

What actually happened?

Quality Gate & Comparative Analytics

What actually happened?

Launch Gate Validation

Base Model Selection

YOLOv8 Nano (n)

YOLOv8 Small (s)

YOLOv8 Medium (m)

Alpamayo-R1-10B

Alpamayo-1.5-10B

Execution Device

NVIDIA DGX Spark Remote Cluster Setup:

Training Controls

What actually happened?

Real-time Output Logs & Curves

Auto-Labeler Parameters

What actually happened?

Interactive Predictions Explorer

Quality Score Summary

What actually happened?

Issue Diagnostics & Recommendations

AI Triage Copilot

Multi-Camera Viewport

NuRec Telemetry & Kinematics

Human Bounding Box Triage

Chain-of-Causation (CoC) Reasoner

Gemma 4 Audit Critique:

Side-by-Side Annotation Boundary Comparators

Key Safety & Quality Metrics

SOTA Annotation Metrics Reference

Export Deployment Engine

Finetuned Weights Details

IDE Model Context Protocol Servers

Raw config file: mcp_config.json

California Statewide Conflict Map

Conflict Analysis Data Grid (FHWA SSAM)

Street Detail

🛠️ Preprocessing Complete: Configure Pipeline Tools

📖 Guidance Panel

📂 Dataset Configuration YAML

🚦 FHWA SSAM Safety Surrogates

⚡ Remote Cluster Training (DGX Spark)

🧭 Full Sensorflow Studio (React)

Raw config file: `mcp_config.json`