CausGround grounds action preconditions for embodied robot planning
Original title: CausGround: Causal Interaction Grounding for Adaptive Embodied Planning
Without parameter updates, reported VirtualHome success rises from 24.9% to 53.3% and RoboCasa from 43.0% to 57.0%.
CausGround is a non-parametric framework that turns interaction experience into reusable action-centered causal knowledge of preconditions and effects. At planning time it grounds those dependencies in current observations for backward and forward reasoning. On adapted benchmarks the authors report higher vision-based planning success across MLLM backbones, and real-robot tests produce executable high-level sequences for sequential manipulation without model updates.
- VirtualHome average task success rises from 24.9% to 53.3%.
- RoboCasa closed-loop planning success rises from 43.0% to 57.0%.
- SWITCH interaction-reasoning accuracy rises from 25.8% to 31.9%.
- Knowledge is stored as object attributes and inter-object relations.
From the original article
1 min read at huggingface.co
Multimodal large language models (MLLMs) show promising reasoning capabilities, but reliable physical interaction requires planning that accounts for when actions are executable and how they change the world. These preconditions and effects are environment-dependent and often not captured by MLLMs, making adaptation challenging.
By Yijun Shen, Diqi He, Jinbang Guo · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.