ODE or Bullet in Gazebo 13.x when reproducibility and accuracy conflict

Answer

If bit‑wise reproducibility across runs is required, ODE remains the safer choice in Gazebo 13.x.

Why ODE is preferred

• ODE’s solver is single‑threaded by default; with a fixed <max_step_size> and <real_time_update_rate> the simulation produces identical trajectories on the same binary/OS.
• Bullet can be forced into a deterministic mode by disabling its parallel solver (<bullet><solver><type>pgs</type><iterations>50</iterations><max_step_size>0.001</max_step_size></solver></bullet>), but floating‑point order differences may still cause tiny variations across different CPUs or compiler builds, so strict reproducibility is not guaranteed.

Steps to configure ODE for reproducible, faster‑than‑real‑time runs

1. Set a fixed physics step, e.g. <max_step_size>0.001</max_step_size> in the world SDF.
2. Match <real_time_update_rate> to the inverse of the step (e.g., 1000 for 0.001 s) or lower if you want fewer updates.
3. Choose a real‑time factor >1 (e.g., <real_time_factor>10</real_time_factor>) – the engine will take the required number of sub‑steps per render step as long as max_step_size remains small enough to keep the solver stable.
4. Keep the world’s <physics><ode> block single‑threaded (default) and avoid enabling <ode><num_threads>.
5. Run a short verification: spawn a known object, record its pose after a fixed simulated time, and compare logs across multiple runs.

Memory behavior with repeated spawn/delete

Both ODE and Bullet release the collision bodies and geometries when a model is deleted. In Gazebo 13.x the dominant source of memory growth over many cycles is fragmentation of the simulator’s internal object pools, not a leak in either physics engine. Monitoring resident memory (e.g., ps -o rss -p <pid>) after a warm‑up of a few thousand cycles will show whether it stabilizes.

When Bullet might still be considered

• If you can tolerate sub‑micro‑second variations and need Bullet’s superior contact handling for high‑poly meshes, enable the single‑threaded solver as shown above and verify reproducibility on your exact hardware/compiler.
• Otherwise, stick with ODE for regression testing.