Safety-Gated Pusher

Problem

Autonomous agents operating in physical or simulated environments must make movement decisions quickly. Without an explicit safety gate in the action planning loop, an agent can execute a planned action that intersects a no-go zone before any constraint check occurs.

The design challenge: make the safety check part of the action-planning loop, not an afterthought — and make the decision and its evidence cryptographically verifiable.

Product concept

A MuJoCo robotic-pushing workflow that checks a planned action against a safety constraint before execution. If the planned path intersects a no-go zone, the action is blocked and logged. If the path is clear, the agent pushes the object toward the goal and records signed evidence.

Created for Robothon 2026. Submission accepted and scored: approximately 76.5/100 observed on the leaderboard.

How the system works

• A robot arm plans a pushing action in a MuJoCo simulation
• The planned path is checked against a no-go zone constraint before execution
• Scenario A: the planned path is clear — the action is allowed and the object moves toward the goal
• Scenario B: the planned path intersects the no-go zone — the action is blocked before execution
• After an allowed execution, passive contact telemetry is recorded (contact count, maximum contact force)
• An Ed25519-signed JSON receipt is generated including the decision, outcome, and telemetry
• Headless EGL rendering works for automated and CI environments

Main conceptual workflow

The diagram below is a conceptual illustration of the system flow. It does not reproduce internal documentation.

My contribution

• Designed and implemented the safety gate logic in the action-planning loop
• Implemented the MuJoCo simulation workflow for the pushing scenario
• Implemented Ed25519 signed evidence generation with contact telemetry
• Validated both the allowed (Scenario A) and blocked (Scenario B) execution paths
• Submitted via public pull request at commit b9761d3 (folder: submissions/safety-gated-pusher/)

Technical implementation

• MuJoCo — robotics simulation environment
• Ed25519 — cryptographic signing for evidence receipts
• Headless EGL rendering — for automated and CI execution

Evidence and validation

• Public pull request visible at the link below (commit b9761d3)
• Submission accepted and scored: approximately 76.5/100 observed on the leaderboard
• Score disclosure: this reflects the score visible at the time of observation — it is not a rank, prize, award, finalist result, or judge endorsement
• Ed25519-signed JSON receipts with contact telemetry (contact count, maximum contact force) included in submission materials

Limitations and transparency

• Simulation only: the workflow operates in a MuJoCo simulation — it has not been deployed on physical robotic hardware
• Observation uncertainty: the score of approximately 76.5/100 was observed on the leaderboard; final scoring details are not confirmed
• No real-world robotics: this does not represent deployment in a production or physical environment
• Contact telemetry reflects simulation physics, not real-world sensor data