CausGroundCausal Interaction Grounding
for Adaptive Embodied Planning

Yijun Shen^1,2Diqi He^2,3Jinbang Guo²Wanyue Zhang^2,4,5Jiaming Mi¹

Huan Yang²Zhongyuan Wang²Zhiwei Yu^2,*Kai Zhang^1,*

¹ East China Normal University² Beijing Academy of Artificial Intelligence (BAAI)

³ Northwestern Polytechnical University

⁴ School of Artificial Intelligence, University of Chinese Academy of Sciences⁵ Institute of Automation, Chinese Academy of Sciences

* Corresponding authors

Paper

Code Video

Reason backward through CausGround.

RoboCasa

Task goals

Conditions & effects

Source record

Retrieval from recorded mechanisms. Initial scene conditions remain to be checked; this animation is not a robot execution.

24.9%53.3%

Average task success

43.0%57.0%

Closed-loop planning success

25.8%31.9%

Interaction reasoning accuracy

Planning performance (%)
Planner	Method	VirtualHome Success rate	VirtualHome Final progress	RoboCasa One-shot	RoboCasa Replanning
Qwen3.5-9B	Vanilla	10.7	23.8	31.0	36.7
	SelfReVision	22.7	34.1	26.8	26.5
	LookPlanGraph	20.0	39.7	19.7	28.6
	CausGround	44.0	61.5	45.1	46.9
Qwen3.5-27B	Vanilla	33.3	46.9	27.9	57.8
	SelfReVision	42.7	55.0	18.3	39.1
	LookPlanGraph	56.0	71.6	24.4	53.1
	CausGround	62.7	79.8	27.9	75.0
Qwen3.5-35B-A3B	Vanilla	30.7	44.4	21.9	34.4
	SelfReVision	37.3	47.9	15.1	4.9
	LookPlanGraph	33.3	47.2	19.2	23.0
	CausGround	53.3	72.6	43.8	49.2

VirtualHome results are macro-averaged over short, medium, and long tasks.

Illustrated causal retrieval with real robot recordings.

Citation details will be added with the paper release.