This paper presents PRISM: an instruction-conditioned refinement method for imitation policies in robotic manipulation. This approach bridges Imitation Learning (IL) and Reinforcement Learning (RL) frameworks into a seamless pipeline, such that an imitation policy on a broad generic task, generated from a set of user-guided demonstrations, can be refined through reinforcement to generate new unseen fine-grain behaviours. The refinement process follows the Eureka paradigm, where reward functions for RL are iteratively generated from an initial natural-language task description. Presented approach, builds on top of this mechanism to adapt a refined IL policy of a generic task to new goal configurations and the introduction of constraints by adding also human feedback correction on intermediate rollouts, enabling policy reusability and therefore data efficiency. Results for a pick-and-place task in a simulated scenario show that proposed method outperforms policies without human feedback, improving robustness on deployment and reducing computational burden.
Figures & tables
Figure 1: PRISM pipeline overview. First a (non-expert) user provides demonstrations on a generic task to generate a generic policy from imitation (left). Under a new task defined by the user in natural-language, this policy is iteratively refined using an LLM that generates reward candidates and improves them according to user feedback (right).
Figure 2: Execution of PRISM pipeline exemplified through evaluation experiment.
Figure 3: Evolution of success metrics for place goal and vertical constraints across four methods. Policies are deemed learned when the total episode reward remains stable for a sustained interval of training steps, marked on both plots with colored dots. For illustration, representative post-training trajectories for PRISM are shown; the EUREKA (RL-only) baseline is excluded from the plot as it stayed to an unsuccessful idle policy.
Imitation learning enables robots to acquire manipulation skills from demonstrations, but the resulting policies can fail outside the training data, while collecting more demonstrations requires substantial human effort. Human-in-the-loop reinforcement learning uses corrective feedback during online training, but typically learns the complete task policy rather than refining a pretrained imitation policy. We introduce Res-HIL, a human-in-the-loop residual reinforcement learning framework that learns corrective actions on top of a frozen imitation policy. Each human intervention provides two complementary learning signals: direct supervision of the residual policy and reward shaping of preceding autonomous behavior. Res-HIL combines these signals with zero initialization of the residual policy to stabilize and accelerate online learning. We evaluate Res-HIL on five contact-rich manipulation tasks spanning high-precision and long-horizon behaviors. With only 20 initial demonstrations, Res-HIL outperforms state-of-the-art full-policy human-in-the-loop reinforcement learning and residual fine-tuning without human guidance on every task after ten minutes of online training. Res-HIL improves its pretrained base policies and outperforms imitation policies trained with five times more demonstrations. An ablation study shows that direct residual supervision is critical to performance, while intervention-aware reward shaping substantially improves training efficiency.
Mariia Iavorskaia, Christian Dietz, Sebastian Albrecht +1
Siemens AG, Research and Predevelopment, Germany. · Technical University of Munich, Germany.
Distilling expert demonstration data into large generative models using behavioral cloning is a scalable approach to learning capable policies for robotic control, particularly for dexterous manipulation. Reinforcement learning (RL) can be used as a means to finetune these policies further using additional experience. An open question is whether RL is more sample-efficient than collecting more human demonstrations. Prior work has finetuned large pretrained policies in a scalable fashion by applying RL to a smaller residual policy that corrects the pretrained model. However, for the typical sparse reward tasks, RL algorithms can struggle to optimize the behavior in a sample-efficient manner. We explore inverse reinforcement learning, where a dense reward function is learned from expert demonstrations, potentially reducing the challenge of RL finetuning. We specifically consider coherent imitation learning, an IRL method that facilitates improvement of the BC policy through using a specific reward formulation with theoretical guarantees. We show that our IRL method maintains or improves the performance of pi-0.5 on all six sparse manipulation tasks and achieves a ≥90% success rate on five out of six complex manipulation tasks, outperforming RL-based baselines using sparse rewards. By ensuring our initial pretrained finetuning policy is optimal for our initial reward and critic, our method circumvents the initial drop commonly seen in RL finetuning and enables faster improvement.
Christian Scherer, Joe Watson, Theo Gruner +3
Technical University of Darmstadt · Zuse School ELIZA · University of Oxford +2
Recent advances in large-scale pretrained vision-language-action models have improved robot policy learning, but directly deploying such policies in user-specific environments remains challenging due to limited generalization, which inevitably requires collecting a dataset tailored to the target environment. Teleoperation yields well-aligned data but is costly and difficult to scale, whereas simulation scales easily but struggles to resemble the target environment and generate task-specific trajectories. To meet both simultaneously, we propose PRISM, an end-to-end pipeline that generates personalized robotic datasets from a single image and a natural-language instruction. PRISM constructs digital cousin scenes that are semantically and geometrically aligned with the user environment yet diverse at the instance level, and synthesizes executable demonstrations without human teleoperation. Extensive experiments show that policies trained on PRISM-generated datasets outperform those trained on baseline-generated datasets on LIBERO and LIBERO-Plus, achieve up to 100% success rate on three real-world manipulation tasks, and maintain stronger performance when evaluated in environments that differ from those seen during training.