Tune Kamino

Kamino is a beta Newton solver selected through a task-exposed newton_kamino preset. This guide isolates and measures a reproducible problem before changing its solver settings. The generated API documentation for KaminoPADMMSolverCfg and KaminoDVISolverCfg remains authoritative for current fields and defaults.

Prerequisites

Read Backends and Presets for preset semantics and Backend and Solver Presets for task integration. Prepare the asset for Newton with Prepare an Asset for Newton with MJWarp. Confirm that the task exposes newton_kamino through its --help output or the environment catalog; do not infer Kamino support from another Newton preset.

Start from an explicit baseline

Run one small visual smoke test before training or benchmarking:

uv run python scripts/environments/zero_agent.py --task Isaac-Cartpole-Direct --num_envs 4 --viz newton physics=newton_kamino

Then fix the initial state, seed, action sequence, and reset distribution. For that reproduction, record penetration, slip, energy, a task metric, a convergence diagnostic, and runtime. Change one decision at a time so every candidate can be compared with the same baseline.

Choose PADMM or DVI

Choose PADMM first for contact-heavy tasks. DVI is a candidate for mechanisms with relatively few active contacts. If a DVI task needs an explicit KaminoDynamicsCfg override, disable preconditioning; otherwise leave the dynamics override unset so Newton selects the compatible defaults.

Validate reset and state consistency

Validate root and joint state writes through the Isaac Lab asset APIs before tuning the solver. Kamino’s maximal-coordinate body poses must agree with the reduced joint state after every reset. Investigate body-pose and joint-state disagreement, including custom reset code and reset overlaps, before changing convergence or stabilization settings.

Choose and size collision handling

Compare Kamino’s internal detector with Newton’s collision pipeline using the same full reset and contact distribution. use_collision_detector on the Kamino solver selects the internal path, which KaminoCollisionDetectorCfg configures; when it is disabled, NewtonCollisionPipelineCfg configures Newton’s path. Size the active path only from observed capacity failures, missing contacts, or warnings, and use those generated APIs for the exact controls. Collision capacity cannot correct invalid geometry or reset penetration.

Tune timestep and substeps

Hold the policy period and fixed reproduction constant. Only after the model and reset are valid, compare a smaller simulation timestep or more num_substeps for the same policy period. Recheck the recorded physical, convergence, task, and runtime metrics; do not change policy timing merely to hide an unstable simulation step.

Tune convergence and stabilization

Tune KaminoPADMMCfg iteration and residual controls only for a reproducible PADMM convergence failure. For DVI, tune KaminoDVICfg alternating iterations and tolerance only for the same kind of evidence. Change KaminoConstraintsCfg stabilization controls only after collision and reset validation. More solver work cannot repair inconsistent state, invalid collision geometry, or reset overlap.

Optimize only after validation

Disable diagnostics before performance measurements. Retain only changes that improve the recorded metrics, and separately measure sparse Jacobian, CUDA graph, collision, substep, and iteration costs when each applies to the selected solver and workload. For the parallel workflow, see Tune MJWarp; it does not replace Kamino-specific validation or tuning.

The diagnose-first order is: validate the baseline, reset state, and collision path; compare timestep and substeps; investigate convergence and stabilization; then optimize measured costs.