Reports
Technical reports, with the code behind them
Full write-ups of the work, from a production driving model's silent failure to a physics-backend comparison for legged-robot RL. Every entry links to a public repository you can rerun, and every entry states its status, including the results that came back null.
Technical reports
Written up in full as technical reports, with PDFs. Not submitted to a venue and not peer reviewed.
-
Silent Collapse: A Distribution-Shift Teardown of a Production Driving Model
Yusuf Guenena · Technical report
Stress-tests openpilot's shipped driving model under simulator-rendered input shift and shows most outputs collapse while its exported uncertainty stays flat, then catches the failure with a recurrent-state monitor.
-
Swapping the Physics Engine Under a Legged-Robot RL Task: A Comparative Analysis of PhysX and Newton in Isaac Lab 3.0
Yusuf Guenena · Technical report
Runs the same Go2 locomotion task on the PhysX and Newton physics backends in Isaac Lab and measures speed (Newton trains 1.64x faster end to end, IQM over 10 seeds on one GPU) and how the dynamics differ. A later config audit found the learning-curve comparison is not backend-only; a rerun is owed.
Work in progress
Active engineering with a public evidence ledger. No manuscript exists for these yet.
-
HELIX: a runtime reliability layer for ROS 2 robots
Yusuf Guenena · Work in progress · no manuscript
A reliability layer that monitors a live ROS 2 robot, diagnoses likely faults and audits recovery commands, run on a Unitree GO2. Physical recovery is the next hardware test.
-
go2-phoenix: a gated sim-to-real pipeline with failure-seeded replay
Yusuf Guenena · Work in progress · no manuscript
A sim-to-real pipeline that trains GO2 locomotion in Isaac Lab and gates every export to the robot. A walking policy has passed every live stand gate on the physical GO2 and taken its first steps on a free floor; it under-tracks the command, and an on-robot probe measured unmodelled joint friction as the cause.
Null and archived results
Studies whose hypothesis did not hold, with their code and data.
-
IPFD: counterfactual replay for failure localization
Yusuf Guenena · Archived, falsified · preregistered study
Tested whether replaying a simulator from a saved state can pinpoint when a robot failure became unrecoverable. A preregistered study showed it cannot.
-
ashfall: failure-driven robot learning, tested in toy studies
Yusuf Guenena · Null result, held
Tested whether failures carry uniquely useful information for fixing robot policies. Five small preregistered studies all came back negative.
The rest of the work
Supporting repositories, with their state, are indexed on the work page.