Prove
LiveProof-of-Work · Oct–Nov: every verified proof counts. 28 days leftAll outcomes
Skills

Robotics 7: Embodied foundation models

6 weeks · 2 milestones

Fine-tune an open vision-language-action policy on your own data. Stage 7 of the Varsity Curriculum: Robotics. A simulation-only route counts at every step; free resources only.

Part of: Robotics · Proves for: Engineers, UG

Milestone map

Milestone map

2 milestones

Fine-tune an open robot policy on your own data

3–5 weeks (about 8 hrs/week, est)

Build an open vision-language-action (or other pretrained robot) policy fine-tuned on demonstrations you recorded yourself, for example with LeRobot, compared with the policy before fine-tuning. Keep it in one public repo and commit as you go. No hardware needed: this milestone is done in simulation.

Proof required

Link the public repo with your dataset card (what you recorded and how), the fine-tuning code and settings, and success rates before and after on the same task.

What gets checked

  • The demonstrations are your own and the dataset card says how they were made.
  • Before and after are tested on the same task and conditions.
  • The base model and its licence are named.

Common mistakes

  • One big upload with no commit history.
  • Results the repo cannot reproduce.

Powstik Guide

A public repo with an open vision-language-action (or other pretrained robot) policy fine-tuned on demonstrations you recorded yourself, for example with LeRobot, compared with the policy before fine-tuning.

Steps

  1. Make a public repo.
  2. Set up the simulator from its free docs.
  3. Build the smallest version that works, then measure it.
  4. Write a README: what it is, how to launch it, what the numbers show.
  5. Ask someone who knows robotics to check it.

Template

README
Stage 7: Embodied foundation models
What is here:
How to launch it:
Results:
Limits:

What gets sent back

  • No measurements.
  • Code that does not work from the README.
  • Copied code with no history.

Resources

Foundationstart here

What a verifier looks for

  • The demonstrations are your own and the dataset card says how they were made.
  • Before and after are tested on the same task and conditions.
  • The base model and its licence are named.
  • Open the repo: does the code produce what the recording shows?

You'll sign in first, then come straight back here.

Give it an instruction someone else writes

1 week (about 4 hrs, est)

Show the fine-tuned policy following a new instruction someone else writes, on hardware or in simulation. Do it in person or on a live video call. Simulation-only route: if you have no hardware, a live run in the simulator counts.

Proof required

Name the person who watched and add the recording or your notes of what happened, including anything that failed.

What gets checked

  • The instruction is new and written by the other person.
  • Success or failure is recorded.
  • Limits of the model are explained.

Common mistakes

  • A pre-recorded run passed off as live.
  • Cutting the failures.

Powstik Guide

A live demo to one person who knows robotics, with notes or a recording.

Steps

  1. Agree the test with the person first.
  2. Let them choose the conditions.
  3. Show it live and record what happened.
  4. Write down what failed and why.

Template

Demo notes
Who watched:
What they asked for:
What happened:
What failed:

What gets sent back

  • The person did not see it live.
  • No notes or recording.

Resources

What a verifier looks for

  • The instruction is new and written by the other person.
  • Success or failure is recorded.
  • Limits of the model are explained.
  • Did you watch it happen live?

You'll sign in first, then come straight back here.