Inspect position-step integration
Enable PositionStepOptions.record_velocity_candidates to record the native
velocity and nominal integration for each iteration of solve_position_step():
options = embodik.PositionStepOptions()
options.record_velocity_candidates = True
result = solver.solve_position_step(q, targets, options)
for candidate in result.velocity_candidates:
print(candidate.step_index, candidate.integration_dt)
print(candidate.solver_velocity, candidate.integrated_velocity)
The option defaults to False. Recording does not change targets, constraints,
integration, or candidate acceptance. Both the single-target and multi-target
overloads support it.
Each PositionStepVelocityCandidate contains:
| Field | Meaning |
|---|---|
step_index |
Zero-based velocity iteration within this call |
integration_dt |
Actual integration interval, including adaptive scaling |
reference_q |
Configuration before this nominal integration |
solver_velocity |
Native velocity before integration clamps and projections |
integrated_velocity |
Manifold difference from reference to candidate divided by the integration interval |
candidate_q |
Nominal configuration before priority, collision, and final acceptance guards |
task_scales |
Scales from this iteration's native velocity solve |
Velocity fields have dimension nv; configurations have dimension nq.
They differ for continuous and floating-base joints. Use the robot's manifold
difference rather than subtracting configurations to reconstruct velocity.
These are candidate records. A later guard may backtrack or reject the
configuration, so the last record need not equal result.q_solution. Task
scales are local to each velocity solve and do not certify final pose tracking.
The records do not include every retry or continuous swept-path safety.
Native call timing
result.computation_time_ms measures the complete native solve_position_step()
call: admission checks, inner iterations, recovery trials and final validation.
Both overloads use the same scope, including early failures. Nested retries are
included once in the outer elapsed time. Python argument conversion, trace-file
I/O, rendering and application work are outside this metric; measure the control
loop separately for those costs.
Earlier versions copied the last velocity solve's time into this field, so historical values can undercount multi-step or recovery calls. Rebaseline timing comparisons across this correction; the accounting change does not establish a runtime improvement.