MODULE 35 / TRAIN & EXTEND
Tails, arms and new policies
Plan an expressive tail or simple arm as a complete robot revision. Learn how reach, mass, actuator choice, collision geometry and control strategy interact before ordering hardware or widening a policy.
By the end, you can…
- Specify a gesture or manipulation task with measurable limits.
- Update the mechanical, power and simulation model for an extension.
- Choose between commanded joints and a wider policy, then define requalification tests.
Completed physical baseline; Modules 7–10, 17, 20 and 31.
A measured robot mass model and a known-working rollback configuration.
Choose a behavior before choosing a motor
A tail that gestures while the robot is parked is a different task from a tail used for dynamic balance. An arm that waves is different from one that carries an object at full reach while walking. Write the intended behavior, reach, payload, speed, duration and permitted robot motion first. These requirements determine loads and the control problem.
Start with one expressive degree of freedom and no carried payload unless the user’s task requires more. Authored trajectories can produce useful gestures without reinforcement learning. Grasping adds end-effector design, perception, contact planning and payload uncertainty; it deserves a later milestone rather than being hidden inside “add arms.”
Added mass changes more than total weight
A component far from the body can shift the center of mass and increase rotational inertia substantially. Moving it also applies reaction loads to the torso. Include motor, gearing, bearings, fasteners, wiring and any object being carried. A visually small mechanism can change the balance problem if its mass is far from the support feet.
Create a simulation variant with explicit body mass, center of mass, inertia, joints and collision geometry. A colored visual mesh alone does not update those properties reliably. Compare the original robot with the added mass fixed in several poses before permitting motion. Fixed-pose comparisons help separate payload effects from control-timing effects.
Select an actuator from the complete load case
For a simple horizontal arm, gravity torque is approximately payload mass × gravity × horizontal distance, plus the arm’s own distributed weight. Motion adds acceleration, impacts, friction and thermal duty. Check the torque-speed curve and continuous conditions, mounting and output-bearing loads. Peak torque alone is not a useful continuous operating promise.
RobStride’s vendor documents provide model-specific manuals and identify heat-sink conditions for rated torque. The existing BDX-R motor family does not automatically make one model suitable for a new mechanism. Also verify voltage, firmware/protocol, feedback, wiring, bus capacity, regulator demand and physical clearances. Do not select a purchased part from a CAD-shaped placeholder.
Choose how the new joints are controlled
One route keeps a separately commanded gesture controller and adapts locomotion to the resulting mass and motion. Another expands the policy’s action and observation contract so it coordinates more joints. Neither is automatically compatible with the old ONNX export. More action slots require corresponding training, joint mapping, scaling and deployment changes.
Requalify in stages: fixed mass, supported mechanism, slow gestures while supported, standing, then walking under measured conditions. Train any new learned behavior in simulation and compare exported outputs before hardware use. Keep the original removable configuration and its policy; a modification is complete when the affected evidence is renewed, not when the animation looks convincing.
WORDS YOU NOW KNOW
- Morphology
- The robot’s body structure and arrangement of joints.
- Lever arm
- The perpendicular distance from a joint axis to a force’s line of action.
- Duty cycle
- How demanding activity is distributed over time.
- Control contract
- The agreed joint ordering, units, timing, limits and meaning of commands and observations.
YOUR CALL
Finished with this module?
0 of 9 lab steps checked. 0 of 3 questions answered. You decide when the module is complete. Revisiting or editing it will not change your decision.
You can reopen it later. Checking every box will not automatically complete it.Go to the source
Primary references for this module. Use the version and hardware revision named in the lesson.
RobStride vendor manuals and specificationsMuJoCo body inertial propertiesBDX-R training tasksModification budget