Source-linked AI summary
CAG-WAM: Causal Action-Gain World Action Models for Robot Manipulation
Peizheng Guo, Cheng Zhang, Xingyu Zhang, Changwen Zheng, Wenwen Qiang
TL;DR
World Action Models can evaluate futures that combine candidate-action effects with environmental evolution, so overall future scores may not characterize action-related change. CAG-WAM uses causal analysis and paired-future evaluation to guide policy learning and action revision; experiments on multiple benchmarks support its effectiveness.
Problem
A candidate’s predicted future can combine action-induced and environment-driven changes, so its overall evaluation may not characterize the change associated with that action.
Source: §1 · p. 1Method
CAG-WAM uses an SCM and compares action-conditioned and environmental futures under the same observation, using their evaluation difference as causal action gain for training and, when needed, action revision.
Source: §1 · pp. 2–3, §6 · p. 10Results
Experiments on multiple benchmarks support the effectiveness of CAG-WAM.
Source: §6 · pp. 9–10Takeaways & Limitations
CAG provides a basis for characterizing candidate-action change relative to environmental evolution and guiding policy learning and action generation.
Source: §1 · p. 3
Abstract
from arXiv · showhide
World Action Models (WAMs) combine action generation with future prediction, allowing predicted futures to support action evaluation and robot decision making. However, the future associated with a candidate action may contain action-induced and environment-driven changes. Directly evaluating the predicted future may not fully characterize the future change associated with the candidate action. For example, a candidate action that worsens a favorable environmental future may still yield a favorable outcome, while one that improves an unfavorable environmental future may still yield an unfavorable outcome. To analyze this issue, we formulate a Structural Causal Model (SCM) for action evaluation in WAMs. The SCM formalizes how action-induced and environment-driven changes jointly shape future evaluation, and reveals a backdoor path through which environmental evolution can also affect the future evaluation associated with a candidate action. Intervention and backdoor adjustment account for this path, providing a causal basis for characterizing the future change associated with each candidate action. Based on this analysis, we propose Causal Action-Gain World Action Model (CAG-WAM). For each candidate action, CAG-WAM predicts an action-conditioned future and an environmental future under the same current observation, and defines their evaluation difference as causal action gain (CAG) for characterizing the future change associated with the candidate action. During training, we incorporate CAG into Group Relative Policy Optimization (GRPO) to guide robot policy learning. During inference, we evaluate the initial action with CAG and, when needed, revise it using predicted future information. Experiments on multiple benchmarks support the effectiveness of CAG-WAM.
1 INTRODUCTION
World Action Models use predicted futures to inform robot action evaluation, but those futures can combine action-induced and environment-driven changes. CAG-WAM uses a causal analysis and paired-future evaluation to characterize each candidate action’s change and guide policy learning and action generation.
1 INTRODUCTION
- Predicted futures can mix action-related and environmental changes, so their overall evaluation may not reveal whether a candidate action improved or worsened the environmental future.The examples include actions that worsen a favorable environmental future yet appear favorable overall, and actions that improve an unfavorable future yet appear unfavorable overall.Source: §1 · pp. 1–2
- The SCM identifies a backdoor path from candidate action to future evaluation through the current observation and environment-driven changes.The current observation influences both action generation and environmental evolution, linking the candidate action and outcome beyond the action-induced changes.Source: §1 · p. 2, Figure 1 · p. 2
- CAG-WAM pairs an action-conditioned future with an environmental future under the same observation and defines their evaluation difference as causal action gain.The same future evaluator scores both futures, making the difference characterize the candidate action relative to environmental evolution.Source: §1 · pp. 2–3
- During training, CAG serves as the GRPO advantage; during inference, low-CAG initial actions can be revised using predicted future information.The revised action is combined with the initial action before execution.Source: §1 · p. 3
- The paper formulates action evaluation causally, proposes CAG-WAM for policy learning and action revision, and reports effectiveness across multiple benchmarks.The contributions include an SCM describing how action-induced and environment-driven changes jointly affect future evaluation.Source: §1 · p. 3
2 RELATED WORK
Prior work combines future prediction and action generation in WAMs, and applies causal reasoning to different robot-learning problems. This paper focuses on causally characterizing candidate-action changes in future evaluation to guide policy learning and action generation.
2 RELATED WORK
- Prior WAMs model action-conditioned futures for candidate evaluation, action selection, or refinement, while this work focuses on causally characterizing action-related change against environmental evolution.The cited examples include unified video-and-action modeling, action-conditioned prediction, tactile-future ranking, and future-based action refinement.Source: §2 · p. 3
- Prior causal robot-learning work addresses state-variable identification, counterfactual demonstrations, visual confounding, and action-dependent future prediction; this work instead analyzes candidate actions and environmental evolution in future evaluation.CAG-WAM uses that analysis to characterize each candidate’s future change and guide policy learning and action generation.Source: §2 · p. 3
3 PROBLEM SETTINGS AND ANALYSIS
The section frames the challenge of evaluating candidate actions when predicted futures include both action-induced and environment-driven changes. It analyzes the issue causally and motivates paired-future evaluation for characterizing action-related change.
Problem Setting
- A WAM receives visual and proprioceptive observations and task instructions, samples candidate actions, predicts future observation sequences, and scores them with a scalar task-related evaluator.The evaluator may return a task reward, value estimate, task-progress score, or a combination.Source: §3 · p. 4
Motivation Analysis
- Overall future evaluation can obscure whether a candidate action improves or worsens the future relative to environmental evolution.Candidate-relative ranking compares sampled actions but does not establish whether an individual action improves or worsens that environmental future.Source: §3 · p. 4
Causal Analysis
- The SCM represents environmental changes as dependent on the current observation and exogenous factors, while candidate actions produce action-induced changes; both shape the future observation.The future observation is subsequently evaluated under the task instruction.Source: §3 · pp. 4–5
- CAG-WAM’s paired-future model scores action-conditioned and environmental futures with the same evaluator, and its framework uses their comparison in training and inference.The resulting CAG is the policy-optimization advantage; at inference, future feedback supports another action generation when CAG falls below a revision threshold.Source: Figure 2 · p. 5, Figure 3 · p. 6
- Because the current observation affects both action generation and environmental evolution, a backdoor path can make action evaluation reflect more than action-induced change.The identified path is At ←Ot →Et+1 →Ot+1 →Yt.Source: §3 · p. 5
- Intervention and backdoor adjustment block that path by fixing the action and replacing the action-dependent observation distribution with p(o).The analysis presents this adjustment as a causal basis for characterizing each candidate action’s future change.Source: §3 · p. 5
4 METHODOLOGY
CAG-WAM compares action-conditioned futures with environmental futures to characterize candidate-specific change, then uses that signal to guide policy learning and inference-time action revision.
4.1 ARCHITECTURE: CAUSAL ACTION-GAIN WORLD ACTION MODEL
- The FC-AM generates candidate action chunks, while the PF-WM predicts action-conditioned and environmental futures for shared evaluation and subsequent action revision.The futures share the current observation and task instruction but differ in their action conditions.Source: §4.1 · p. 6
4.2 TRAINING: CAUSAL ACTION GAIN-GUIDED POLICY OPTIMIZATION
- Training computes each candidate's causal action gain by comparing its predicted future with the environmental future, then incorporates that gain into GRPO.The paired futures use the same current observation and task instruction, providing a matched comparison.Source: §4.2 · p. 6, §4.2.1 · p. 7
4.2.1 CAUSAL ACTION GAIN
- CAG compares evaluations of action-conditioned and environmental futures from the same observation; its sign and magnitude indicate the relative evaluation difference.Positive gain means the action-conditioned future is evaluated higher, and negative gain means it is evaluated lower.Source: §4.2.1 · p. 7
4.2.2 CAUSAL ACTION GAIN-GUIDED OPTIMIZATION
- The policy uses each candidate's causal action gain as its advantage in clipped GRPO, tending to encourage positive-gain actions and suppress negative-gain actions.GRPO optimizes grouped policy samples without requiring an additional critic model.Source: §4.2.2 · p. 7
4.3 INFERENCE: CAUSAL ACTION GAIN-GUIDED FUTURE REVISION
- At inference, an initial action below the revision threshold triggers another action generation conditioned on feedback from the two predicted futures.The feedback uses the difference between the action-conditioned and environmental future representations.Source: §4.3 · pp. 7–8
- The revised action is combined with the initial action before execution, with λ controlling the revised action's contribution.Future information is incorporated selectively according to causal action gain.Source: §4.3 · p. 8
5 EXPERIMENTS
Experiments evaluate CAG-WAM on RoboCasa and LIBERO, with results across both benchmarks and a case study of action revision after a slip.
5.1 EXPERIMENTAL SETUP
- Evaluation uses task success rate on RoboCasa’s seven skill families and LIBERO’s four suites, comparing CAG-WAM with representative robot-policy methods.The listed baselines include Diffusion Policy, Octo, OpenVLA, and Cosmos Policy-Predict2.5.Source: §5.1 · p. 9
5.2 MAIN RESULTS
- 74.2% average success rate across 24 RoboCasa tasks; 99.8% average success rate across the four LIBERO suites.LIBERO success reached 100% on Object and Goal, with high performance on Spatial and Long.Source: §5.2 · p. 9
5.3 ANALYSIS
- After an intermediate slip causes an inaccurate grasp, future revision adjusts the subsequent action and recovers to complete the task.Without revision, execution misses the target.Source: §5.3 · p. 9, Figure 4 · p. 9
6 CONCLUSION
The conclusion frames WAM action evaluation as a causal problem and presents CAG-WAM as a way to characterize candidate-action effects relative to environmental evolution.
6 CONCLUSION
- Action-induced and environment-driven changes jointly shape future evaluation, so overall future evaluation may not characterize a candidate action’s associated change.CAG-WAM compares action-conditioned and environmental futures, defining their evaluation difference as causal action gain.Source: §6 · pp. 9–10
- CAG enters GRPO during training, while inference evaluates an initial action and revises it when needed using predicted future information.The conclusion reports that experiments support CAG-WAM’s effectiveness.Source: §6 · p. 10
AI USE STATEMENT
The authors used generative AI for grammar and language polishing, not research ideation, methodology, analysis, or substantive writing.
AI USE STATEMENT
- Generative AI was used for grammar and language polishing, but not for research ideation, methodology, analysis, or substantive writing.The authors take full responsibility for the paper’s contents.Source: AI USE STATEMENT · p. 10