GRPO post-training
Connected complete robot attempts to shared GRPO policy updates.
A round plan sets tasks, contexts and configurable budgets for two collectors. Simulation collects grouped rollout windows. A physical robot waits for an operator Start and records complete attempts. An admission gate passes only valid, scored, version-matched and group-matched data to one shared GRPO learner with a single algorithm state, running from dataset to trainer. The learner publishes a versioned policy, and the next round starts after workers adopt it. The learner also saves and resumes from checkpoints that carry round and training statistics.
Technical details
Robot attempts vary in length and need consistent grouping for GRPO updates. I connected complete-attempt collection to the shared training core, with leave-one-out advantages, EMA baselines and per-task return normalization. Collection quotas, optimizer-step boundaries and algorithm state are saved for continuation.
