Inspiration
VIDYUT means electricity. Electricity does not care whether a machine is a drone, rover, humanoid, or spacecraft; it flows through every connection and exposes every weakness. That became the idea behind VIDYUT: follow an autonomous machine through its geometry, electronics, controller interfaces, environment, and failure response so teams can find integration problems before a costly physical trial.
Today, these layers are usually split across CAD tools, wiring documents, simulators, firmware terminals, and manually assembled reports. A robot can look correct in CAD and its controller can work independently, yet the complete system can still fail because of an invalid pin mapping, insufficient power, a duplicated bus address, or an unsafe response to sensor loss. Smaller robotics and space teams may not have the budget, equipment, or safe locations needed to discover those problems through repeated physical testing.
What it does
VIDYUT turns that fragmented process into one reviewable workflow:
- Describe the machine. Import geometry and physical properties, then declare components, power rails, communication buses, controller interfaces, pin mappings, and engineering assumptions in a versioned machine manifest.
- Validate readiness. Detect issues such as incompatible voltage ranges, insufficient current capacity, missing power, duplicate I2C addresses, reused pins, invalid interfaces, and incomplete mechanical properties before execution.
- Compose repeatable failures. Select an environment and build a timed test plan from 25 templates covering sensing, navigation, actuation, power, communications, compute, mission, and combined faults.
- Execute in SIL or connect a controller through HIL. The built-in deterministic preview runs immediately. The Web Serial HIL bridge supports a guarded controller handshake, bounded commands, sequence checks, a watchdog, and an emergency stop for stationary bench testing.
- Capture evidence. Export the machine fingerprint, scenario seed, fault timing, assertions, telemetry, controller I/O, and per-test results as JSON, CSV, or a self-contained HTML report.
Our hackathon demo uses a humanoid scenario with IMU bias and joint-torque loss. VIDYUT visualizes truth versus controller state, measures stability and joint load, and produces repeatable evidence from the same configuration and seed. A one-click judge path also demonstrates a drone under GNSS drift, command-link loss, and wind.
The humanoid's complete autonomy software is not programmed yet. However, VIDYUT's HIL path can connect to its physical NUCLEO controller and validate communication, bounded actuator commands, safety limits, and hardware responses before the complete autonomy stack is ready.
How we built it
VIDYUT is a browser-based developer tool built with JavaScript, Node.js, Three.js, WebGL, and the Web Serial API. Its core data contracts are vidyut.machine.v2 for machine definitions and vidyut.evidence.v2 for reproducible results.
The importer supports VIDYUT JSON/YAML packages plus engineering data from URDF, SDF, STEP metadata, glTF/GLB, and KiCad netlists. A deterministic, seed-based executor models aerial, ground, legged, and spacecraft behaviours. The Three.js proving ground provides articulated machines, lit terrain, shadows, multiple cameras, route tracking, and platform-specific motion constraints. Automated tests verify all four reference machine families, all 25 test templates, deterministic replay, readiness failures, file parsing, HIL safety behaviour, and evidence generation.
The architecture deliberately separates test intent from the execution backend. VIDYUT's preview is runnable today, while its exported adapter package defines integration boundaries for PX4/Gazebo, FMI 3.0 co-simulation, and ASAM OpenSCENARIO concepts. We do not claim that the preview is a certified physics simulator.
How we used Codex and GPT-5.6
We used GPT-5.6 through Codex as an engineering collaborator throughout the build, not only to generate a first draft. Codex helped us:
- turn the initial idea into an implementable architecture and versioned schemas;
- research simulation, HIL, component-data, and scenario standards;
- implement the import pipeline, electrical validator, deterministic test engine, Three.js proving ground, Web Serial bridge, evidence reports, and automated tests;
- diagnose hardware integration issues around the NUCLEO-G474RE, PCA9685, MG996R servos, I2C mapping, and separate logic/actuator power domains;
- review claims and expose unavailable external backends honestly instead of showing simulated success.
GPT-5.6 also converts engineering intent into bounded, structured scenario configurations and can draft component records from manufacturer information. Those AI-generated fields remain unconfirmed until an engineer reviews them. Deterministic code—not the language model—executes the scenario, calculates the metrics, and decides each assertion result.
Challenges we faced
The biggest challenge was deciding where AI should stop. A language model is useful for interpreting intent and organizing incomplete engineering information, but it should not invent physical properties or act as an authoritative physics engine. We therefore made provenance, confirmation state, deterministic execution, and honest backend availability part of the product design.
Hardware brought a second challenge. High-current servos cannot be powered from a controller's logic rail, and a browser connection is not automatically a safe HIL system. We separated logic and actuator power, required a common signal ground, and designed the bridge around bounded commands, acknowledgements, watchdog behaviour, and a latched emergency stop.
We also had to create one coherent experience without pretending to replace established tools such as Gazebo, PX4, Isaac Sim, or MuJoCo. VIDYUT's role is the readiness, test-intent, adapter, reproducibility, and evidence layer that connects those ecosystems.
Accomplishments that we are proud of
- A complete Machine Setup -> Scenario Builder -> Run & Evidence product flow.
- Readiness validation that can block an unsafe electrical configuration before execution.
- Exactly 25 reusable failure-test templates and four reference machine families.
- Deterministic replay with fingerprints, seeds, thresholds, telemetry, and per-test results.
- A real Web Serial controller bridge with safety-oriented protocol checks.
- Exportable JSON, CSV, HTML, coverage, and backend-adapter packages.
- Eighteen passing automated tests covering the core workflow.
- A deployed judge-ready demo that works without requiring a local rebuild.
What we learned
We learned that the most valuable digital-twin problem is not only making a convincing animation. Engineering teams need traceable inputs, explicit assumptions, repeatable failure timing, measurable assertions, and evidence they can compare between software, controller, and physical trials. We also learned that AI is strongest when it helps engineers structure and investigate a problem while deterministic systems retain responsibility for execution and measurement.
What's next for BAC VIDYUT
VIDYUT is not only a hackathon project for us. We intend to develop it into a startup serving robotics, physical-AI, autonomy, and space-engineering teams. After recognizing the scale of this problem, we began exploring the process of applying for up to INR 50 lakh in business financing to support product development, engineering infrastructure, hardware integration, and early customer pilots. This financing has not yet been approved, but starting the process reflects our commitment to continue building VIDYUT after the hackathon.
Log in or sign up for Devpost to join the conversation.