roboticsfactbearishDeploying reinforcement learning on real robots remains constrained by high interaction costs, since each physical rollout is costly and reflects only one realized action-outcome pathArtificial Intelligence27 Jul 2026http://arxiv.org/abs/2607.02431v1