Skip to content

Robot dynamics and task execution

A reachable pose can still require too much torque. A collision-free path can still be impossible to track. A gripper can reach its commanded pose while the object stays on the table. Learn to expose these gaps with small calculations and runnable models before adding a learned policy.

Prerequisites: Robotics Foundations, especially frames, kinematics, control, and planning; Python functions and lists. The optimal-control lesson introduces its matrix notation. Layers: L1 workload and task contract, with L3–L5 consequences for scheduling, control execution, and compute budgets. Roles: robotics, robot learning, embedded control, and physical-AI systems engineering.

Access: Python 3.10+ and a CPU; standard library only. No simulator, GPU, external model, or robot is required. The calculations describe specified mathematical fixtures. Hardware control and contact simulation are later stages.

What to learn and build

Lesson New idea beyond the foundations Artifact Study time
1. Torque, gravity, and actuator limits Motion depends on inertia and available effort Torque/current calculation and three controller traces 60–90 min
2. LQR and constrained predictive control Optimize a model, then distinguish costs from hard constraints Riccati check, feasible plan, infeasible case, and failed settling result 90–120 min
3. From a path to executable motion Check swept geometry and time the motion Collision counterexample, duration calculation, tracking-error test 60–90 min
4. From a grasp pose to object success Evaluate contacts and object motion, not just the hand Transform calculation, force assumptions, and paired task outcomes 60–90 min
Lab and independent transfer Change one model assumption and explain the result Reproducible JSON, tests, prediction, and one failed case 1–2 h

Entry diagnostic

Attempt these before reading. Each identifies a particular gap rather than a reason to restart the roadmap.

  1. A 2 kg point mass is 0.5 m from a horizontal joint. With g=9.81 m/s², what holding torque is needed? Can an 8 Nm actuator supply it?
  2. A controller minimizes position error and effort. Does that objective prevent crossing a wall or exceeding a motor limit?
  3. Two trajectory endpoints are free of obstacles. Is the connecting motion free?
  4. The hand rises 20 cm. What must you observe to establish that a pick succeeded?
Answers
  1. 2 × 9.81 × 0.5 = 9.81 Nm; 8 Nm is insufficient under this model.
  2. No. A penalty and a hard constraint are different. Inspect the actual optimization problem and applied command.
  3. No. The robot sweeps a region between endpoints. It can intersect an obstacle even when both endpoints are clear.
  4. The object's lift and its relation to the hand, plus the task's retention and placement conditions. A hand trajectory alone is insufficient.

Run the reference

From the repository root; commands were tested here with Python 3.12.3 on Linux on 2026-10-18:

COURSE='Phase 5 - Advanced Topics and Specialization/Track D - Robotics/Robot Dynamics and Task Execution'
python3 "$COURSE/labs/control_lab.py" all --output /tmp/robot-task-execution.json
python3 -m unittest discover -s "$COURSE/labs" -p 'test_*.py' -v

The supplied 24 tests check hand calculations, a finite-difference energy derivative, an independent two-step control objective, the Riccati value identity, interior collision extrema, timing peaks, transforms, and object-level failure. They finish in seconds. The JSON includes implementation hashes and Python version; retain it with your command and repository revision.

Some reference outcomes deliberately fail: an undersized actuator cannot hold the target; a coarse action grid keeps moving near the goal; a hand can rise without the object. Passing the exercise means predicting and detecting these outcomes, not changing the success definition until every row is green.

Continue into a complete system

Use ROS 2 interfaces to carry the command, feedback, and timing contracts. Use the Isaac engineering sequence to test an actual scene, moving robot geometry, and contact dynamics. Add robot learning when a learned component addresses a demonstrated limitation.

Complete this course when: another learner can reproduce your baseline and variation, verify the worked numbers, identify the first failed assumption, and explain what requires a simulator or physical measurement. Apply the shared rubric to your evidence.