SELF-PACED WORKSHOP
Organize tasks and accumulate skills with ASPIRE
Let a model propose plans, call skills, inspect feedback and revise programs; distinguish ASPIRE, GR00T and Agnes Harness.
A reasoning model needs executable robot interfaces
“Go around the chair and send the ball into the goal” combines goal interpretation, subtask organization and motion. A model can propose and revise a plan, while controllers maintain balance and handle contact. A complete plan may require both a high-level sequence and a geometric route; prose alone does not supply either validated execution or collision avoidance.
Scroll horizontally to view the full diagram
ASPIRE: turn attempts into reusable experience
ASPIRE uses code-as-policy: an agent generates control programs, inspects multimodal execution traces, diagnoses failures, revises code and retains validated experience in a skill library. This learning primarily updates programs and skill knowledge, rather than necessarily updating neural weights. The public October 2026 v1.0 workflow includes iterative debugging and evolutionary search.
Scroll horizontally to view the full diagram
For example, missing goal-post bases on a pale floor can prompt inspection of images and failure records, followed by another viewpoint or a revised perception interface. A program exiting normally is not a goal, and redefining scoring to pass a trial is not an improvement. The reusable output is a tested method.
Adapting the framework to Microduck
| Interface | Reusable foundation | Remaining integration |
|---|---|---|
| Observation | RGB, ToF, IMU and joint records | Consistent timing, frames and observation age |
| Action | Walking, skating, kicking and stopping | Bounded calls, timeouts, cancellation and mode semantics |
| Outcome | Contact, crossing, fall and recovery checks | Diagnostic feedback plus independent scoring |
| Skill library | Existing control and failure knowledge | Validate new programs on unseen starts/scenes |
Official examples target LIBERO-Pro, Robosuite, BEHAVIOR-1K and documented real-arm systems. There is no ready Microduck adapter for this course. Renaming arm joints does not turn end-effector control into biped control. Reuse ASPIRE’s exploration and skill-library approach while implementing body-motion and football interfaces for Newton.
How ASPIRE and GR00T can cooperate
Start by letting ASPIRE call existing motor skills to study task organization. Once a Microduck GR00T adaptation is validated, learned skills can replace selected action components. GR00T maps current observations to actions; ASPIRE organizes goals, calls, checks and revisions. This is a proposed integration, not an officially delivered Microduck stack.
Stable skill interfaces and retained experience reduce repeated task programming. The robot still needs the physical capability for a task: a new prompt alone cannot guarantee pickup without suitable grasping hardware. New demonstrations, skills or tools may still be necessary.
Extension: where Agnes Harness fits
| Aspect | ASPIRE | Agnes Harness |
|---|---|---|
| Focus | Refine robot programs and skills through execution | Connect models, tools, task state and devices |
| Reuse | Validated robotic skill libraries | Skills, sessions and execution records |
| Current scope | Robotic exploration/research workflows | Developer-preview general agent harness |
| Course role | Main task-agent methodology | Optional device-integration comparison |
They share concepts but are not equivalent products. Agnes’s MHS device integration and plugins merit study, but the inspected public main branch contains no ready Microduck adapter. We keep it as an extension rather than repeat the agent loop in another full chapter.
The linked Microduck community demo is a useful reference for questions about action tools, completion checks and recovery. Its page was unavailable in this review environment; its implementation and autonomy level have not been independently verified or counted as course results.
Feasible with an adapter and independent evaluation
The method and open interfaces are promising. A Microduck execution bridge, skill library and autonomous exploration trials remain to be implemented. Begin with observe, bounded movement, stop and check, then extend to football. Separate development and held-out seeds, bound trial budgets and record success, recovery and model-call costs.