Exploring the Limits of Combinatorial Generalization in Embodied Manipulation

Exploring the Limits of Combinatorial Generalization in Embodied Manipulation. By Xupeng Zhang and four others, from Shanghai Jiao Tong University. Published on the eleventh of October. Its code repository is worked out: about one hundred and seventy source files, with a readme and a licence.

In embodied manipulation, Vision Language Action models can execute atomic skills. Long-horizon manipulation asks for more: recombining those skills without having demonstrations for every possible combination. The common route, full instruction training, treats chaining skills as long-horizon imitation. A controlled study shows where that leads. Trained on a limited set of skill combination sequences, full instruction reaches forty percent ordered success on the combinations it covered, but only seven point seven percent on recombination. The authors read that gap as memorization of sequence templates rather than robust recombination of skills. Telling those two apart calls for a test built around recombination itself, one where no combined trajectory is ever trained on.

RoboCombine is that test. It is a benchmark for skill recombination without combined skill training trajectories. It scores two things: ordered success, and transition conversion across the boundaries between subtasks. The second metric looks at exactly the moment one skill hands over to the next.

Alongside it sits Hierarchical Subtask Execution, or HSE, a diagnostic framework. HSE puts a shared modular interface over three components: a planner, a trigger, and a policy. That interface makes skill recombination visible, and lets a long-horizon failure be decomposed into separable modes instead of being counted as a single miss. With the pieces separated, the question becomes which of them actually breaks.

The first result is that atomic competence alone does not ensure ordered execution. A policy can perform each skill and still fail the sequence. Failures concentrate at transitions. A downstream subtask has to launch from the handoff state its predecessor leaves behind, not from the canonical start it saw as an atomic task. The authors name this mismatch the start state gap.

To address it they introduce Pose Robust Task Initialization, or PRTI, a lightweight post-training strategy for atomic tasks that broadens start state coverage beyond canonical starts. Under oracle-timed HSE on RoboCombine, PRTI reaches eighteen point one six percent ordered success. Random perturbation reaches four point two one percent, and a robot-reset oracle reaches nine point four seven percent. None of the three uses combined skill training trajectories.

The conclusion is about where the difficulty sits. Decomposing a long task into subtasks the robot already knows is not enough on its own. Combinatorial generalization needs atomic skills that are robust to handoffs, so that progress in the first stage converts into reliable execution downstream. The weak point is the seam between skills, and that is where the benchmark, the diagnostic and the fix all aim.