CausGround: Causal Interaction Grounding for Adaptive Embodied Planning

TL;DR

CausGround is a non-parametric framework that distills interaction experience into a reusable library of action preconditions and effects, organized in a factorized world-state representation of object attributes and inter-object relations. At planning time, task-relevant causal dependencies are grounded against the current observation to support backward reasoning over preconditions and forward reasoning over effects for action selection and multi-step planning on top of a frozen MLLM. Reports VirtualHome avg task success 24.9 → 53.3%, RoboCasa closed-loop planning 43.0 → 57.0%, SWITCH interaction-reasoning accuracy 25.8 → 31.9%, and real-robot sequential manipulation from MLLM-interpreted goals — all with no parameter updates to the underlying MLLM.

Key claims

Interaction experience can be distilled into a reusable, structured library of action preconditions and effects that generalizes across environments without MLLM retraining [Abstract].
A factorized world-state representation (object attributes + inter-object relations) is the substrate the preconditions and effects index against; task-relevant causal dependencies are grounded against the current observation at planning time [Abstract].
The planner supports both backward reasoning over preconditions (what must hold before an action is executable) and forward reasoning over effects (how the world changes) for action selection and multi-step planning [Abstract].
On adapted VirtualHome, CausGround lifts average task success from 24.9% to 53.3% across MLLM backbones [Abstract].
On adapted RoboCasa, closed-loop planning success rises from 43.0% to 57.0% [Abstract].
On SWITCH, average interaction-reasoning accuracy rises from 25.8% to 31.9%, with the authors attributing the gain to improved action selection, recovery, and verification [Abstract].
Real-world experiments show the system reasons backward from MLLM-interpreted goals to produce executable high-level action sequences for sequential manipulation [Abstract].

Method

CausGround sits as a non-parametric layer around a frozen multimodal LLM. From interaction experience (collected across simulated and real environments), it identifies the preconditions and effects of actions and stores them, keyed against a factorized world-state representation that tracks object attributes and inter-object relations. At planning time, given a goal, the planner (i) retrieves task-relevant causal dependencies from the library, (ii) grounds them in the current observation (so preconditions become checks on the current state and effects become predictions over successor states), and (iii) uses backward chaining over preconditions and forward simulation over effects to select the next action and compose multi-step plans. Because the knowledge store is external and symbolic rather than baked into weights, new environments are addressed by extending the library rather than retraining the MLLM.

Evaluation spans three regimes: adapted VirtualHome and RoboCasa for high-level planning (open-loop and closed-loop), SWITCH for interaction reasoning, and a real-world sequential-manipulation deployment driven backward from an MLLM-interpreted goal.

Results

VirtualHome high-level planning: 24.9 → 53.3% average task success across MLLM backbones.
RoboCasa closed-loop planning: 43.0 → 57.0% success.
SWITCH interaction reasoning: 25.8 → 31.9% average accuracy; authors attribute the lift to action selection, recovery, and verification.
Real-world: executable high-level action sequences for sequential manipulation, derived by backward reasoning from an MLLM-interpreted goal; no quantitative aggregate reported in the abstract.
All reported gains achieved with no updates to MLLM parameters.

Why it’s interesting

CausGround is a clean embodied-regime instance of the external experience library + frozen base model pattern that Training-Free Group Relative Policy Optimization stakes out for text tool-use agents: both move the “policy update” out of the base weights and into a structured, inspectable artifact that reduces required inference-time work. Compared to Reconstruct, Practice, Go Real: Guided Self-Improvement for Embodied Agents RPG — the current closest neighbor on the VLA Models “no-weight-update, library-is-the-policy” row — CausGround keeps the artifact causal and relational (preconditions over attributes/relations with backward chaining) rather than a growing procedural skill library, which is the first filed instance of that specific flavor in the wiki. The MLLM-interpreted-goal → backward-chain → executable action sequence pipeline is also the embodied counterpart to the Playful Agentic Robot Learning RATs / Embodied Turing Machines: Stateful Code for Robot Recursive Self-Improvement COAP line, but with the knowledge representation shaped for planning over procedures.