GoPi: Hybrid Open/Closed-loop testing and RCA pipeline for AV stacks

[mk3039]

Hello Apollo Community,
On behalf of the team at Energesque, we're excited to share our project with you

Overview

Energesque is initiating a pilot program for GoPi, a closed-loop testing and root cause analysis (RCA) pipeline for AV stacks. Pilot participants are granted full license access to the product suite (RCA, Crash, and AHA engines) in exchange for agreeing to serve as a reference account. GoPi interfaces with a stack's native messaging protocol, executing without internal code access or runtime modifications, with out-of-the-box integration for stacks built on open-source platforms (such as Apollo).

Architecture & Workflow

1. Scene Mutation (Crash Generation)

The system ingests baseline MCAP logs and mutates camera, LiDAR, radar, odometry, and GNSS topics to create a mutated sensor fusion that simulates a collision. This pipeline produces synthetic collisions mapped to four standardized profiles: Head-On (HO), Side-Swipe (SS), T-Bone (TB), and Rear-End (RE).

2. Audit Harness Analysis (AHA) Engine

Executes a lightweight, hybrid open/closed-loop simulation based on an MCAP record by injecting mutated sensor data into the target stack, then capturing and executing the vehicle’s control commands to update a kinematic model. This allows users to quickly record and benchmark a stack’s response based on records of both real and simulated crash incidents, rather than requiring a full scenario rebuild.

3. Root Cause Analysis (RCA) Engine

Post-execution, the RCA engine parses MCAP logs using a rules-based diagnostic engine to extract key system telemetry, identify failure patterns, and classify collision events. A local LLM is deployed as a natural language presentation layer to summarize and format the outputs of the rules engine.

This approach allows users to quickly identify points of failure in critical edge/crash cases from both real or simulated logs.

Performance Baseline

We have utilized this pipeline to measure the crash mitigation performance of the Autoware, Apollo, and Openpilot platforms against the HO, SS, TB, and RE crash profiles.

apollo-brain-hands_NuScenes-v1.0-mini-scene-0061-head-on_rca_report.json
autoware-full_NuScenes-v1.0-mini-scene-0061-head-on_rca_report.json
openpilot-full_NuScenes-v1.0-mini-scene-0061-head-on_rca_report.json

Integration Requirements

GoPi deploys stack-specific middleware adapters to interface directly with baseline configurations without requiring target-side source code modifications.

• Autoware: ROS 2 topic injection.
• Apollo: CyberRT bridge via dedicated channel mapping.
• Openpilot: Cereal/ZMQ message serialization interface.

Custom/Closed Source Stacks: Requires an Interface Control Document (ICD) to map proprietary messaging protocols to GoPi's unified data schema.

Team

GoPi was developed by a team with backgrounds in autonomous vehicles, robotics, communications, software architecture, and defense systems. Core contributors include Homayoun Yousefi'zadeh (Adjunct Prof. UC Irvine) and Hamid Jafarkhani (Chancellor Prof. UC Irvine), a Software Architecture Lead from the Defense Industry with expertise in Autonomy and Avionics.

The co-founders have a 20-year history of joint project execution, having previously completed and secured federal and state agency funding for several prior software projects. Energesque is a graduate of the highly competitive UC Berkeley SkyDeck accelerator program.

Contact Us

If your team is interested in utilizing GoPi to benchmark safety margins on your stack, contact us at pilots@energesque.com. We will provide the crash library and sample RCA reports to demonstrate the output format prior to deployment.