Ticket in. Deployed robot out.
File a ticket for a robot task. The foundry runs a baseline, buys exactly the data that fixes the failures from people on their phones, fine-tunes π0.5 and ships it.
On the line now
ConnectingLoading the live numbers.
How the line works
- Baseline
- Diagnose
- Plan data
- Collect
- Review
- Train
- Eval
- Ship
- 01Baseline
The robot runs the task 10 times on its current policy, so the ticket starts from a measured number.
- 02Diagnose
Claude watches the failed trials and groups them by why they failed.
- 03Plan data
The planner decides what data fixes those failures, how much, and what it costs inside the budget.
- 04Collect
Contributors claim paid tasks on their phones: teleop demos, takeover recoveries and resets.
- 05Review
Every episode is checked for gaps, jerk and success before anyone is paid.
- 06Train
π0.5 is fine-tuned with LoRA on one H100 on Modal. About 50 minutes per run.
- 07Eval
The new checkpoint runs 10 trials. The ticket shows the rate with its 95% interval.
- 08Ship
If the rate clears the target, the checkpoint goes to the robot. If not, the line loops back.
Two ways in
Contributors
Open the link on your phone, claim a paid task, drive the arm from the browser or reset the scene. Payouts run through Stripe in test mode. You need the 6-character code shown next to the QR.
Operators
Plug in a robot, file a ticket with a target success rate and a budget, then watch it move down the line. Operator login uses the operator passcode, not the contributor code.
Built with
- SpacetimeDB · state, event log, live UI
- Factory Droids · long-running agents
- Modal · training and inference
- Claude · diagnosis and planning
- π0.5 · the policy
- I2RT YAM · the arm
A Cal Hacks build. Money on this site is Stripe test mode.