Marc ik stel voor om het volgende indefinit cycle te proberen te maken met de SO100s
We moeten die modellen namelijk naar production krijgen en ik denk dat als dit lukt we facking goed gepositioneerd zijn om het met moelijkere taken te proberen
Task Demo — SO-100 Block Stacking
A robot that stacks three coloured blocks in order, then resets the environment itself, and repeats indefinitely.
The Problem
Brittle Models
VLA models trained on narrow datasets break on minor shifts in lighting, object placement, or surface friction.
Silent Failures
The robot completes a motion but the task outcome is wrong. Nothing alerts. Nothing retries. We need a way to detect when the world is not in the state the model expected.
Data Gaps
The model has never seen the edge cases that matter most in production. We need a method to identify those gaps and fill them efficiently.
No Recovery
Once in a failed state the robot has no strategy to return to a known-good configuration. We need to understand what recovery looks like and how to train for it.
Open Questions
How much data do we need?
This is asking: what is the minimum number of demonstrations required to reliably fine-tune a model on a new task, and what factors push that number up or down.
How do we detect a failure?
This is asking: what does a failed state actually look like in data, what signals are reliable indicators, and how do we build software that acts on them in real time.
How do we recover?
This is asking: once a failure is detected, what is the best path back to a state the robot can continue from, and how do we train a policy specifically for that transition.
How do we add data efficiently?
This is asking: how do we incorporate new demonstrations into an existing model without degrading what it already knows, and what is the right method for doing that incrementally.
General remarks
Ik denk dat we de vragen door ze zo op te splitsen lekker hapbaar hebben gemaakt ook denk ik dat how do we detect een failure mischien een determisische manier heeft om dit op te lossen