Isaac engineering lab sequence¶
Use this sequence after the introductory lectures. It adds deliberate debugging, task tests, and deployment evidence to the existing simulator and learning material.
Prerequisites: Python; vectors, coordinate transforms, and basic feedback control; Lectures 01–04 before scene work, 09–10 before training. Read the shared learning guide for study rhythm and grading.
Effort: approximately 20–30 hours for the exercises below, excluding simulator installation and long training jobs. Begin with 16 environments, short runs, and state observations; scale only after correctness.
Version and hardware contract¶
Reference documentation and selected source contracts rechecked 2026-10-18; use the compatibility record for current 6.2 versus the named 6.1/EA baseline: Isaac Sim 6.1 requirements and Isaac Lab v3.0.0-EA installation. Use a resolved Isaac Lab commit and its lockfile, simulator version/image digest, driver, task/asset revision, physics backend, and renderer. A release branch can change after this course was written.
Full Isaac Sim workflows need a compatible RTX system. An 8 GB card is an experimental reduced path, below the documented 16 GB minimum, not a supported baseline for all labs. Kit-less training is a separate path with different features; passing it does not validate Isaac Sim rendering or ROS integration.
The existing Lecture 01 contains setup notes. Preserve your installation record and reconcile any version-specific command against the pinned source's help. Do not combine a 2.x tutorial's API with a 3.x task just because their class names look similar.
Lab ladder¶
| Lecture | Main exercise | Evidence | Reduced path |
|---|---|---|---|
| 14 — Reproducible simulation setup | Pin, launch, zero/random-action smoke tests | Manifest and startup/steady-state logs | Inspect metadata and design smoke cases |
| 15 — Frames, units, and assets | Robot/scene contract and transform probes | Frame diagram and round-trip tests | Pure-Python quaternion fixture |
| 16 — Task and reset correctness | Partial resets, rewards, done flags, action mapping | Task test matrix | Reset-isolation fixture |
| 17 — Learning and evaluation design | Train baseline, held-out evaluation, ablation | Learning curves and outcome counts | Wilson-interval calculation |
| 18 — Sensor and simulation performance | Profile physics/render/policy and environment count | Throughput/memory curves | Workload and buffer arithmetic |
| 19 — ROS 2 timing and policy export | Observe-to-act timing and stale-command handling | Timing trace and policy contract | Offline command-age fixture |
| 20 — System identification and transfer | Perturb dynamics and delays, identify failure envelope | Paired evaluation and bounded assumptions | Offline response-data analysis plan |
| 21 — Language-guided robotics capstone | Asynchronous language planner plus bounded robot skills | Trace, rejection tests, closed-loop results | Recorded-event replay |
CPU exercises supplied here¶
The expanded CPU simulation lab adds actual Newton physics, OpenUSD composition, six real ROS replay outcomes, a separate-process TorchScript adapter fixture, scalar system identification, and paired outcome analysis. Use its measured/raw results and explicit optional dependency instructions before running a full simulator.
From this course directory:
The implementation and tests need Python 3.10+ and no simulator. They cover coordinate convention, partial-reset isolation, rate arithmetic, confidence intervals, and command freshness. Their outputs are synthetic checks, not Isaac performance or robot validation.
Expected rate fixture: 16 environments, 100 policy steps, decimation 2, physics dt 1/120 s, 2 seconds wall time gives 800 transitions/s, 100 vector physics steps/s, and per-environment real-time factor 0.833333…. Aggregate transitions/s is not the clock speed of one robot.
Simulator execution route¶
Inside the pinned Isaac Lab checkout, consult the versioned task generator or existing course Lecture 10. Select the actual task ID printed by that checkout. The following upstream-documented command pattern uses a shell variable, not a literal placeholder task:
# Set TASK_ID to a task registered in your pinned project.
: "${TASK_ID:?Set the registered task ID}"
uv run isaaclab zero_agent --task "$TASK_ID" --num_envs 16
uv run isaaclab random_agent --task "$TASK_ID" --num_envs 16
uv run isaaclab benchmark runtime --task "$TASK_ID" --num_envs 16 --num_steps 1000
Record the selected backend and renderer from resolved configuration; defaults may select a kit-less backend. For the Isaac Sim path, use the pinned installation's documented extra and physics preset. First prove scene/reset correctness, then short training, then full runs. These simulator commands were sourced from upstream documentation and were not executed in this repository's CPU-only validation.
Grading¶
A simulation milestone needs a launch log, task contract, negative test, raw measurements, and a reproducible configuration. A policy milestone additionally needs held-out episodes and a baseline. A transfer milestone requires evidence from the target environment; a simulation-only result remains simulation-only.