CausGround: Causal Interaction Grounding for Adaptive Embodied Planning

CausGround distills interaction trajectories into reusable action-centered causal knowledge over a factorized world state, then grounds task-relevant preconditions and effects to guide frozen MLLM planners without parameter updates.

MLLMs show reasoning capabilities, but physical interaction requires planning that accounts for when actions are executable and how they change the world, and these environment-dependent preconditions and effects are not fully captured by MLLMs. Existing adaptation approaches either update model parameters through costly environment-specific training or rely on unstructured interaction experience that is difficult to ground and reuse. Existing representations also do not always jointly capture fine-grained object attributes and inter-object relations required to determine whether an action's preconditions are satisfied.

CausGround is a non-parametric framework that adapts embodied planners without updating model parameters by distilling interaction experience into action-centered causal knowledge. It defines a factorized world state S_t=(O_t,R_t), where each object has a semantic category and attribute-value pairs, and relations are typed edges between objects. From action-level trajectories collected under controlled variations of pre-action factors, it retains state factors that distinguish successful executions from failures as preconditions, retains consistent post-action changes as effects, and discards incidental factors. Extracted preconditions and effects are organized into generic action knowledge K_gen(a) and instance-specific knowledge K_spec(a(x)), and effects that satisfy another action's preconditions create causal dependencies. At planning time, CausGround interprets the task into a goal state, performs backward reasoning to obtain K_causal, retrieves K_sem, forms K_task = K_causal ∪ K_sem, grounds it in the current world state, and a frozen planner selects actions from the observation, state, goal, and task knowledge.

• On VirtualHome, CausGround improves the reported average success rate and average final progress over Vanilla. §4.2
• On RoboCasa, CausGround improves one-shot planning success and closed-loop replanning success over Vanilla when averaged across backbones. §4.2
• On SWITCH interaction reasoning, CausGround improves average accuracy over Vanilla across the evaluated MLLM backbones. §4.3
• Across six models from three families, CausGround improves averaged task success and final progress over Vanilla. §4.5
• In component ablations, full CausGround reports higher success rate and progress than Vanilla and the partial variants. §4.4
• In retrieval ablations with Qwen3.5-9B, semantic retrieval alone improves over Vanilla, causal retrieval yields larger gains, and combining both gives the best result. §D.2

• Left open: controlled interaction experiments were used to construct the VirtualHome library, while RoboCasa and SWITCH knowledge was extracted from repeated official trajectories under a consistency criterion (C.3.1); the paper does not evaluate controlled-intervention construction in those environments.
• Left open: real-world evaluation is qualitative and covers two tabletop tasks, the microwave task and the multi-object task; no quantitative real-robot success rate is reported (E).
• Evaluation limitation: RoboCasa results are reported on the shared subset of cases with no UNKNOWN transition, and UNKNOWN actions are marked separately rather than as ordinary planning failures (A.2.3); behavior on unresolved branches is not measured.
• Evaluation dependency: VirtualHome uses RGB observations repaired with Qwen-Image-Edit-2511 and validated against simulator state (A.1.3); results depend on this repaired benchmark, and no separate no-repair comparison is reported.

• CausGround is non-parametric and does not update planner parameters; it supplies reusable causal knowledge as external planning context. §Abstract
• The world state is factorized as S_t=(O_t,R_t), with object attributes and explicit typed inter-object relations sharing the same schema as preconditions and effects. §3.1
• Action-centered causal knowledge stores preconditions C(a(x)) and effects E(a(x)) extracted from action-level trajectories, discarding incidental factors. §3.2
• Knowledge is split into generic K_gen(a) and instance-specific K_spec(a(x)) regularities across grounded instances. §3.2
• Effect-precondition links compose single-action knowledge into multi-step causal dependencies. §3.2
• Task-grounded planning builds K_task = K_causal ∪ K_sem once per task, grounds it in the current world state, and the frozen planner outputs one action per step. §3.3
• VirtualHome and RoboCasa evaluate visual high-level planning, while SWITCH evaluates Action, Recovery, Verification State, and Verification Action categories. §4.1
• VirtualHome observations inconsistent with simulator state were repaired with Qwen-Image-Edit-2511 and validated before evaluation. §A.1.3