33 written lessons, one arm, and a research question. A project-first path from LLM performance engineering to robot-learning evidence that survives a lab meeting — classical control through VLA deployment, taught for the systems engineer you already are.
The WidowX AI on your desk is a sealed production system: six integrated brushless joints on an internal CAN FD bus behind a 500 Hz real-time controller. The AK60-6 on your bench is the same actuator species with the lid off. Open it up — outrunner, planetary gear, encoder, FOC firmware — and learn the impedance law every modern joint runs before you build policies on top of it.
Open lesson →Units are the written lessons. A phase counts only when its exit criterion is evidenced on the robot — not when the reading is done. Both live in this browser.
Open curriculum →Every phase pairs written lessons with a robot studio, and ends in an artifact plus an exit criterion. External courses are references, not the path.
Your performance-engineering instincts transfer directly here: timestamps, contracts, reproducibility, and failure visibility are the first robotics skills to operationalize.
Read for hardware abstractions, position versus torque control, sensor roles, and the relationship between a robot description and a running station.
CoreSpatial transforms, forward kinematics, and Jacobians (opens in a new tab)MIT Robotic Manipulation · Chapter 3Work through pose composition, kinematic trees, differential kinematics, and constrained differential IK. Skip basic Python review.
SelectedTrossen AI arm in Isaac Sim and Isaac Lab (opens in a new tab)Trossen Robotics · NVIDIA Isaac integrationOptional setup lane: validate the official single-arm USD and IK demos, then inspect the Isaac Lab reach, lift, and drawer tasks. Use simulation as a test harness—not as a substitute for the real arm.
SelectedROS 2 timing and execution essentials (opens in a new tab)ROS 2 official conceptsBuild a just-in-time systems lane covering executors, QoS, tf2, rosbag2, clock domains, hardware timestamps, and tracing. This is the prerequisite for meaningful NITROS work later.
Isaac Sim is useful here for URDF/sensor validation and software-in-the-loop checks. Do not pause real-hardware bring-up to build a photorealistic scene.
A second engineer can reproduce the setup and recover a synchronized episode using only your repository.
Russ Tedrake · living notes
Primary source for the physical manipulation stack.
Russ Tedrake · living notes
Selective depth for dynamics, optimal control, feedback, and contact.
Deep Reinforcement Learning
Primary learning course because its current syllabus reaches robot VLAs.
Sergey Levine · Deep RL
Extra derivation and homework depth, not a second full course.
Read just in time. Complete only the assignments that produce code, data, or intuition needed by the current robot experiment.
NVIDIA's official three-computer pattern separates DGX-class training, RTX/Omniverse simulation, and Jetson runtime. This course maps Isaac Lab across learning and simulation, and Isaac ROS onto runtime—only when each tool solves a measured problem.
Open the NVIDIA robotics playlist (opens in a new tab)Simulation + synthetic data
URDF/sensor validation, software- or hardware-in-the-loop testing, collision scenes, controlled synthetic data.
Do not optimize scene realism before the real task, sensors, and metrics are frozen.
GPU-scale robot learning
Parallel RL/IL environments, domain randomization, vision-in-the-loop training, and scalable policy evaluation.
Use after a real-robot baseline exists; sim success is not the capstone result.
Accelerated ROS 2 runtime
GPU-accelerated perception, image processing, transport, collision checking, and trajectory-optimization nodes.
Adopt only after tracing the ROS graph proves transport or a GEM-replaceable node is limiting the loop.
On-robot edge compute
Measure power-, thermal-, memory-, and sensor-I/O constraints in a deployable edge configuration.
A workstation is the faster starting point. Buy or port only when edge deployment answers a research question.
Generalist policy platform
Cross-embodiment policy experiments and an end-to-end alternative spanning teleop, training, evaluation, and ROS deployment.
Keep OpenPI primary for the WidowX capstone; use GR00T later as a controlled comparison.
World models + data generation
Generate or transform physical-AI video/data for a specific long-tail generalization hypothesis.
Not a prerequisite. Add it only when the baseline reveals a data bottleneck it can plausibly address.
How do inference latency, jitter, action horizon, and replanning policy affect success, throughput, recovery, and motion smoothness in real-world VLA manipulation?
An adaptive scheduler using predicted inference time and disagreement between active and newly generated chunks will improve throughput and perturbation recovery without increasing unsafe acceleration or action discontinuity.
Freeze the dataset, task rubric, held-out split, and primary metric before tuning the final method. Randomize condition order, standardize resets, preserve failures, and report negative results.
Your existing ML background lets you compress lectures. It does not let you compress contact with the real system.
Only material tied to the active experiment
Implementation, collection, deployment, trials
Two structured paper notes
Plots, statistics, failure bookkeeping
Decisions, negative results, public writing