Pick which robot training save to retrain from
Himanshu Kumar's agent picks where a simulated robot retrains from using a graph of past saves and test results; a graph rule changed its pick, and a regressed result was blocked.
Done with Custom agent
What they did
The developer stored every saved version of a simulated robot driving policy in a connected database. Each save was linked to its parent, its test results, its failures and past decisions. An agent uses query tools to find the best save to retrain from. It records why it chose before any training runs. Each save is tested on four suites, 180 test runs in total.
How it went
He says the agent first repeated a past failure even after reading about it. So he made the graph enforce a rule against it. In the final run that rule changed the agent's pick. The retrained robot still regressed, and the gate stopped it from shipping.
Worth knowing
Telling the agent about past failures was not enough to stop it repeating them; the rule had to be enforced by the system. The author also treats a version family with no recorded problems as unexamined, not safe.
Try it yourself
Here is my training history: [list of saved model versions, their parent versions, and test results per test suite]. Find which versions passed every test suite under these limits: [pass thresholds]. Flag version families where a child got worse than its parent. Recommend which version I should retrain from and why. List every option you considered with its rank. Write your reasoning down before I start any training, and wait for my approval.Swap the highlighted [parts] for your own details.
Source: DEV Community · Himanshu Kumar · Oct 17, 2026
Written with AI assistance from the original post and checked against our editorial guidelines. Results are as reported by the source, not independently verified. Spotted a mistake? Tell us.
