- Rating
- 3
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
CADIS is a planning layer for frozen robot policies operating under resource constraints. It uses a learned joint skill model to predict future physical outcomes and resource requirements, compiles resource trajectories into four-parameter capped affine contracts, and composes these contracts to assess feasibility and choose recovery targets. Candidate task-and-recovery programs are screened using model rollouts; the selected program is then checked using independent rollouts and a Clopper-Pearson risk bound before execution with resource monitoring. The paper evaluates accounting accuracy, controlled multistage recovery, and resource-augmented LIBERO-Long manipulation with a fixed policy.
Critical strengths
-
A clear separation between resource accounting and learned prediction. The four-parameter interface preserves minimum entry requirements and exit-resource behavior while accounting for payment order and capacity clipping under the stated model. Its composition and backward requirement calculations are explicit, and Section 4.2 reports agreement with an exact chronological oracle on 3.64 million queries. This is a useful representational contribution that makes the deterministic resource calculation easier to compose and inspect.
-
The selected program is validated using fresh samples. Section 3.4 separates candidate selection from independent validation, addressing the optimism introduced by reusing selection samples. Proposition 1 gives a finite-sample false-certification bound under the fitted rollout law. This is a sensible statistical design, and the paper explicitly states its conditional scope.
Critical weaknesses
- The practical motivation and additional value over established resource planning are not sufficiently convincing. The problem is structurally close to resource-constrained scheduling: tasks consume resources, recovery restores them, and the planner trades progress against waiting. Within the assumed class of capped affine functions, composition and backward resource requirements are obtained primarily through function composition, intersection of feasible domains, and solving inequalities. These derivations are relatively straightforward; the paper needs to justify more clearly their theoretical advance and significance relative to established resource analysis methods, including the energy algebras cited in Sections 2 and 3.2. Regarding the practical motivation, Appendix D uses the action-dependent consumption law prescribed in Equation (21), a fixed replenishment rate, and an interlock rule that stops execution when resources are insufficient. These settings can test resource planning mechanisms, but the paper does not sufficiently explain how these parameters and triggering conditions correspond to actual device limits, or what practical consequences protective stops would have in real operation. Section 3.1 and Appendix A.2 provide a first-order thermal model that offers a possible physical interpretation; however, the evaluation remains limited to simulations and numerical experiments, without validating the correspondence between the main benchmark's resource rules and real device limits. I therefore still find it difficult to judge the practical importance of the problem represented by these controlled settings and the additional value of the proposed combination in real applications.
- Predictor reliability lacks direct validation, and some baseline comparisons are insufficiently specified. Appendix D reports the MLP architecture and training setup, but does not quantify prediction errors in resource demands or interlock-risk calibration on held-out executions, especially underestimation that could lead to resource insufficiency and interlock stops. Proposition 1 and the 512 independent validation rollouts control statistical error under the fitted model distribution; they do not address systematic prediction error. The baseline comparisons are also inadequately specified: Empirical in Figure 2 and Previous CaDiS, Strong Supervisor, and Recovery repair; fixed menu in Figure 4 lack clear implementation and configuration details. Appendix D describes fixed and reactive recovery and adaptations of CBF and Recovery-RL, but does not sufficiently explain their tuning or which conditions are matched across methods. The implementation details, configurations, and matched conditions for some baselines are insufficiently clear, making it difficult to assess the fairness of the comparisons and which differences between methods account for the reported gains. Empirical appears to match or exceed CADIS on some Figure 2 metrics, and the paper does not explain how to interpret this comparison. In the cohort of 12 initial conditions, Figure 2g shows marked differences across fitted seeds in where utility rises along the risk threshold axis and by how much. This raises concerns about the consistency of the utility response to risk threshold adjustment across seeds. The aggregate results in Table 1 at a fixed risk threshold do not explain this pattern.
Minors Comments
- The AI Use Statement describes assistance with drafting and editing the manuscript. If generative AI was also used to create or edit Figure 1, please clarify that use in the disclosure.
Questions for authors
-
Practical significance and novelty: With reference to the thermal model in Section 3.1 and Appendix A.2, please explain which actual device limits the main benchmark's resource consumption coefficients, replenishment rate, and interlock triggering conditions represent, and what practical consequences protective stops would have. What is the additional contribution over established resource planning and the cited energy algebras, and what specific difficulty does learned prediction address in the chosen setting?
-
Stability of risk threshold adjustment: Figure 2g shows marked differences across fitted seeds in where utility rises along the risk threshold axis and by how much, raising concerns about the consistency of the response to risk threshold adjustment. Could the authors explain these differences using the predictions and admission rates for each seed?