Answers are grounded in this paper and cite their sources.
Learning from reward functions or imitating human demonstrations requires human specification for every behaviour, a problem for long-lived self-adaptive systems whose goals and conditions cannot be fully anticipated at design time. Goal-Conditioned Supervised Learning (GCSL) offers a self-supervised alternative: agents imitate their own experience under hindsight-relabelled goals. However, learning only from self-generated data can reinforce an agent's initial biases, and relabelling focuses solely on successes, so the agent cannot learn from its mistakes. We propose Goal-Conditioned Supervised Learning with Negative Feedback (GCSL-NF), which evaluates each trajectory twice: positively with respect to relabelled goals, and correctively with respect to the original goal. The corrective target comes from a similarity function learned contrastively from trajectory-induced neighbourhood relations, so no reward function or geometric distance needs to be specified. Experiments show that GCSL-NF overcomes initial biases, benefits increasingly from negative feedback as learning progresses, and matches or surpasses leading GCSL- and HER-based methods on discrete and continuous goal-reaching tasks. Its learning signal remains informative under corrupted priors, runtime goal shifts, sensor and actuation noise, and constraints enforced only during training. By reducing reliance on prespecified rewards, GCSL-NF suits self-adaptive systems whose objectives are diverse, changing, or hard to engineer.
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: