Skip to content

Size capacities

Step

Does
sets Steel's constraint-row capacities so no row is ever dropped, with headroom, from measured demand.
Passes when
every row family reports zero overflowing world-steps and check_constraint_capacity() passes on full episodes at the target world count.

Inputs. A scene that runs its whole task; the world count you will use; several seeds. Tools. shelf trial for examples; in your own code, SolverSteel(..., row_watermark=True), solver.constraint_row_watermarks() and solver.check_constraint_capacity().

  1. Start generous: raise articulation_constraint_max, free_body_constraint_max (and propagation_constraint_max under the propagation response) well above what you expect.
  2. Run the complete task, through every load (impact, grasp, pour, settling), with row_watermark=True, over several seeds:
    cd examples && uv run python -m shelf trial EXAMPLE --out outputs/EXAMPLE_cap
    
    Read facts.episode.row_watermarks in trial.json: per family the requested high-water mark (*_raw_high_water), *_overflow_world_steps and dropped contact rows; and the contact buffer’s high-water mark against its size.
  3. Set each family to 1.3x its requested high-water mark from a run that did not overflow. A family that overflowed under-reports its demand: double it and measure again.
  4. Re-run at the world count you will ship. The busiest world sets the requirement and it varies between runs, so repeat near the boundary.

Rules. The families and the contact buffer are independent; raising one never repairs another. Any overflowing world-step fails the run however the video looks. Smaller capacities save memory but rarely change speed.

Outputs. The capacities, the high-water marks they came from, and the runs that confirm zero overflow.

Source: agents/steps/size-capacities/README.md, the copy agents read.