BAAI introduces CausGround, a frozen-model causal planner for robot tasks

AIBAAI presents CausGround, which uses simulation-distilled causal knowledge to guide real dual-arm robots without retraining. In the system, a frozen MLLM interprets natural language into a goal state, and CausGround reasons backward through a causal graph to generate valid action sequences, such as Open(Microwave) ➔ Pick(Bread) ➔ Place(Bread, Plate).

AIBAAI, ECNU, NPU, UCAS and CASIA introduce CausGround, a plug-and-play causal interaction grounding framework that the post says doubles embodied planning with zero parameter updates. CausGround adds a lightweight symbolic layer that distills raw interactions into reusable Action Preconditions and Effects, letting agents reason about what is required before an action and how the world changes after it. It works as an external engine for frozen Qwen3.5, InternVL3.5 and Gemma 4 backbones, with no fine-tuning.

AICausGround represents visual inputs as explicit object attributes and inter-object relations, discarding irrelevant scene detail. It plans with bidirectional causal reasoning, identifying preconditions backward from the goal and anticipating state changes forward before acting, which BAAI says avoids repetitive deadlocks and trial-and-error.

AIBAAI reports that CausGround raises average task success from 23.3% to 50.7% across six open-weight backbones, including Qwen3.5 and Gemma 4, without updating any weights. The post says the gains hold in simulated households and real-world benchmarks, and that CausGround keeps performance more robust than baseline planners on 11–22 step tasks.