Learning Object-Centric Spatial Reasoning for Sequential Manipulation in Cluttered Environments

A team of researchers has introduced a new paradigm for robotic manipulation in cluttered environments with their framework, Unveiler. The work, led by Chrisantus Eze, Ryan C. Julian, and Christopher Crick, argues against monolithic, end-to-end AI models for this complex task. Instead, Unveiler explicitly decouples high-level spatial reasoning from low-level action execution. Its core is a transformer-based Spatial Relationship Encoder (SRE) that sequentially identifies the most critical obstacle to remove, a decision then passed to a separate Action Decoder. This modular approach directly addresses the data inefficiency and lack of interpretability often found in large-scale models, positioning Unveiler as a more specialized and effective solution for automation in warehouses, homes, and industrial settings.

The technical breakthrough lies in Unveiler's performance and efficiency. The SRE is trained in two stages: imitation learning from heuristic demonstrations followed by PPO fine-tuning, allowing it to discover superior strategies. In simulations, it achieved a remarkable 97.6% success rate in partially occluded scenarios and 90.0% in fully occluded ones, outperforming both classic policies and modern large-model baselines. Crucially, the system demonstrates impressive generalization: the SRE's spatial reasoning transfers zero-shot to real-world scenes, and the full system was validated on a physical robot requiring only geometric calibration, with no retraining of learned components. This suggests a path toward more robust, sample-efficient, and deployable robotic assistants capable of complex sequential tasks like decluttering shelves or organizing bins.