acceptodds
Under review as a conference paper at ICLR 2027

When the body computes: The information geometry of cheap control

Abstract

A body can sustain useful motion that a controller would otherwise have to produce. But when does exploiting that motion make control cheaper, and can an information-based reward encourage it? We develop an account of cheap control that separates what a policy encodes from what it physically imposes. A common divergence geometry connects two complementary views of embodied dynamics: action influence through empowerment, and state propagation through morphological computation. Rewarding the latter induces policy compression gated by the body, targeting precisely the controller information that the body's own dynamics render redundant. A causal decomposition extends the result to feedback. We then separate informational from physical cheapness. Physical savings depend on an alignment between what the body sustains, what the task permits, and what control costs. Experiments recover the selective compression mechanism across linear and nonlinear control. The resulting theory explains how embodiment simplifies control, and when information savings become physical ones.

Primary area:
reinforcement learning
Keywords:
morphological computation, cheap control, information geometry, policy compression, empowerment, embodied intelligence, information bottleneck, directed information, reinforcement learning, information theory

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_76ttOpenReview ↗
Rating
2
Confidence
1
Informativeness
3
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates how the passive dynamics of a physical system can reduce the need for active control. The authors analyze the relationships among Morphological Computation (MC), Policy Compression, and Control Intervention from an information-theoretic perspective. They derive conditions under which MC-based regularization selectively reduces redundant controller information and when such information reduction can translate into lower physical control costs. Experiments are conducted in a custom linear control environment and the DMC Ball in Cup task.

Critical strengths
  1. Meaningful research motivation. Exploiting a system's passive dynamics to reduce unnecessary control effort is an interesting research problem. The paper also makes a useful conceptual distinction between reducing controller information and reducing actual physical intervention.
Critical weaknesses
  1. Limited experimental validation. The experiments are primarily conducted in relatively simple control environments. While the linear control setting demonstrates reductions in control intervention, the improvements in learned policies are more limited. In the DMC Ball in Cup task, no statistically significant reduction in physical control cost is demonstrated.

  2. Insufficient validation of the theoretical mechanism in more complex environments. The paper demonstrates selective information compression through MC regularization in a linear control setting. However, the nonlinear Ball in Cup experiment optimizes a passive-dynamics-based information overlap objective (BODY) rather than the original MC objective. Moreover, this experiment does not demonstrate a significant physical control benefit. Therefore, the current empirical evidence remains insufficient to establish whether the proposed theoretical mechanism translates into practical improvements in more complex control tasks.

Questions for authors
  1. Could the authors evaluate the original MC regularization objective directly in more nonlinear control environments to further validate the proposed theoretical mechanism?

  2. In the DMC Ball in Cup experiment, why does the observed selective information compression not translate into lower physical control costs?

5 comments
AuthorsOpenReview ↗

Direct nonlinear MC tests, robustness, and the information–effort mechanism

Thank you for recognizing the information–intervention distinction and for requesting direct, replicated nonlinear tests of the original MC objective.

Summary: we now optimize original MC in nonlinear control across ten seeds, with 39% [38,40] less intervention than Compression at 10% task degradation in the pendulum; Ball in Cup now has 22 seeds. The common author comment maps the full revision.

1. Direct tests of the original MC objective.

In pendulum release, gravity supports motion toward the target, pivot damping helps settling, and a separate alignment error still needs active correction. Policies learned with the original behavioral MC objective use 39% [38,40] less intervention than Compression at 10% task degradation across ten independent training seeds, with the advantage in every seed. At 2% degradation, the saving is 18% [16,20] across nine paired seeds. Brackets here and below denote 95% bootstrap intervals. MC removes more body-visible information while retaining more complementary information (Table 1; Appendix B.5, Table 4).

Changing to a freer pivot preserves this selectivity but removes the resolved intervention advantage at 2%. The same task thus tests both parts of the theory: which information MC compresses, and how the body's settling dynamics determine whether that compression yields physical savings.

In the gantry crane, the trolley can coast while its suspended load needs anti-sway correction. Full MC saves 9% [6,11] relative to Compression at 10% degradation across ten seeds; BODY, which isolates the overlap penalty, saves 37% [36,38] (Figure 2; Appendix B.7, Table 9). These experiments directly optimize behavioral MC as well as BODY. Nonlinear information quantities use the declared Gaussian estimates, supported by independent checks in the appendix; intervention in the primary additive noise models is evaluated as transition KL.

2. Replicated Ball in Cup results and the missing physical advantage.

The expanded 22 seed neural policy study includes the original MC objective alongside BODY, Compression, and Direct (Table 3). BODY lowers overlap relative to Compression in 20 of 22 seeds and retains more complementary information in all 22 at the 2% return tolerance. An independent estimator supports the direction of selectivity (Table 1; Appendix B.3). MC also shows no resolved intervention reduction relative to continued task-only training; the revised experiment identifies why.

Approximately 97% of transition KL intervention occurs after the catch, where a nearly constant holding command carries little state information. Information penalties have no direct incentive to remove this expenditure. Direct does, reducing intervention by 96.4% [96.0,96.8] relative to continued task-only training (Table 3). This is a concrete instance of the paper's central distinction: simplifying the state dependence of feedback does not necessarily remove the costly part of actuation.

3. Mechanism validation and the soundness assessment.

The concerns stated in the review concern empirical coverage, replication, and substituting BODY for MC in the original nonlinear experiment. The new direct MC experiments and replicated neural policy study address each of these points.

The theory makes quantitative, testable predictions. Theorem 4 explains why information reduction alone cannot guarantee physical savings. Theorem 5 predicts the direction and leading magnitude of intervention changes near a regular task optimum. The revised quantitative test spans nine pendulum bodies and five objectives. The computed alignment predicts intervention increases in all four negative alignment cases, which the experiments confirm; it also predicts the leading magnitude of intervention changes near the task optimum (Table 5; Figure 7). Proposition 11 and the joint stop experiments additionally address failure of the recoverability condition. Humanoid balance extends the empirical information–effort distinction to articulated contact dynamics (Appendix B.10).

The revision now supports the theoretical account with direct nonlinear MC gains, replicated selectivity, and quantitative tests of when information savings become physical ones. Given these additions addressing the stated basis of the concern, we respectfully request reconsideration of the soundness assessment and overall rating.

AuthorsOpenReview ↗

Follow up on direct MC validation

Dear Reviewer 76tt,

Following your request, we directly tested the original MC objective in nonlinear control across ten seeds, obtaining 39% less intervention than Compression in pendulum at 10% task degradation. The revision also includes 22 seed Ball in Cup results, an explanation of its information–effort gap, and quantitative alignment tests. We have addressed all your concerns and would appreciate reconsideration of your overall rating in light of this evidence. We remain available for any further questions.

AuthorsOpenReview ↗

Followup on direct MC validation and robustness

Dear Reviewer 76tt,

With the discussion period about to end in a day and half, could you please indicate whether the direct nonlinear tests of the original MC objective and the expanded 22 seed Ball in Cup study address your concerns about validation and robustness?

We would greatly appreciate your updated soundness assessment and overall recommendation, together with any specific remaining question that we can address before the window closes. We remain available to respond promptly.

Reviewer_JzjuOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how passive body dynamics can reduce the information required by a controller and potentially lower physical control costs. The authors develop an information-theoretic framework connecting morphological computation, policy compression, and control intervention. The proposed analysis is supported by theoretical results and experiments in linear and nonlinear control environments.

Critical strengths
  • The paper investigates an interesting problem related to embodied intelligence and efficient control.

  • The theoretical framework provides an interesting interpretation of the relationship between passive dynamics and controller information.

  • The Push and Coast experiment demonstrates that the proposed information regularization can reduce control intervention compared with standard policy compression.

Critical weaknesses
  • The strongest results come from a relatively simple, specially designed linear system. It is unclear whether the conclusions generalize to more complex control tasks, like humanoid control.

  • In the nonlinear Ball in Cup experiment, the proposed approach does not demonstrate a significant reduction in physical control cost compared with standard policy compression.

  • The theoretical analysis relies on several assumptions, including specific properties of the dynamics and state distributions. It is unclear how the proposed framework could be applied to practical robotic control problems.

  • The nonlinear experiment uses only one policy-search random seed, making it difficult to assess the consistency of the results.

Questions for authors
  • Can the authors provide stronger experimental evidence in more complex nonlinear environments?

  • Why does the proposed method fail to significantly reduce physical control costs in Ball in Cup?

  • How could the proposed framework be used in practical robotic control, where the passive dynamics may not be known accurately?

5 comments
AuthorsOpenReview ↗

Replicated nonlinear savings, humanoid control, and learned passive dynamics

Thank you for the specific requests for stronger nonlinear evidence, multiple seeds, and practical relevance. The revision now includes substantial learned nonlinear savings, a 22 seed Ball in Cup study, and the requested humanoid experiment. The common author comment maps all additions; the evidence addressing your concerns is below.

1. Stronger evidence beyond the designed linear example.

In pendulum release, gravity moves the pendulum toward its hanging target; pivot damping determines whether it settles or needs active braking, while a separate alignment error still requires correction. Learned MC uses 39% [38,40] less intervention than Compression at 10% task degradation across ten seeds, and 18% [16,20] less at the tighter 2% target across nine paired seeds. Brackets here and below denote 95% bootstrap intervals. MC reduces body-visible information while retaining more complementary information (Table 1; Appendix B.5, Table 4).

In the gantry crane, momentum can sustain trolley travel while the load still needs anti-sway control. BODY relaxes trolley feedback while preserving stabilization, saving 37% [36,38] relative to Compression at 10% degradation across ten seeds; full MC saves 9% [6,11] (Figure 2; Appendix B.7, Table 9). These are learned nonlinear control results with substantial physical gains.

For DMC humanoid balance, we added ten neural controller adaptations per body on the official and a stiffer body. On the stiff body, BODY and Compression reduce their information measures by 58–65%, yet corrective effort increases. Direct reduces that effort by 10.6% [10.3,10.8], with no falls (Appendix B.10, Table 12; Figure 11). The controllers adapt around competent standing behavior. The result extends the central distinction to contact rich control: preserving posture does not eliminate the need for costly corrective feedback.

2. Why does Ball in Cup show selectivity without a physical advantage?

The new 22 seed neural policy comparison resolves the single seed concern. At 2% return tolerance, BODY lowers body-visible information relative to Compression in 20 of 22 seeds and retains more complementary information in all 22. An independent estimator supports the direction of selectivity (Table 1; Appendix B.3).

The physical explanation is now explicit: approximately 97% of transition KL intervention occurs after the catch. The ball has already been caught, but a nearly constant holding command remains. Such a command can be physically expensive while carrying little state information, so information penalties do not directly target that expenditure. Direct does, reducing intervention by 96.4% [96.0,96.8] relative to continued task-only training (Table 3). This gives a concrete mechanism for the information–intervention distinction in Theorem 4. Pendulum and crane establish the complementary case where selective compression does produce physical savings.

3. Which assumptions matter, and how can this help robotics with unknown dynamics?

The revision distinguishes the general identities from their specializations. Equation (5) follows from the mutual information chain rule and the Markov relation A−S−YA-S-Y, without a Gaussian assumption. Recoverability is needed for the simpler MC identity; a new Proposition 11 gives the correction for general actuation. Gaussian assumptions give explicit directional weights, and the physical alignment guarantee is local to a regular task optimum. Nonlinear information estimates are approximate, with independent estimator checks reported in the appendix.

We now test the quantitative alignment prediction across nine pendulum bodies and five objectives: near the optimum, it predicts both savings and the negative alignment cases where intervention increases (Table 5; Figure 7). Held torque and joint stop variants further test action dependent mechanics (Table 6).

For unknown dynamics, using a learned passive model to compute the regularizers closely reproduces the intervention results obtained with known dynamics; task costs and trajectories are evaluated in the original plant (Table 7). This supports a practical route: learn what the mechanics sustain from operating data, then use task–body–action alignment to identify which feedback can be relaxed within an acceptable performance loss.

The revision therefore addresses the linear-only, single seed, humanoid, and unknown model concerns with new evidence tied to the theory. We would appreciate reconsideration of the overall assessment in light of these additions.

AuthorsOpenReview ↗

Follow up on nonlinear evidence, robustness, and practical relevance

Dear Reviewer Jzju,

We would appreciate your feedback on the revised evidence. We addressed all your concerns with replicated nonlinear savings (39% for MC in pendulum; 37% for BODY in crane, relative to Compression at 10% task degradation), 22 seed Ball in Cup results explaining the information–effort gap, humanoid experiments, and a learned passive model check. Given these additions, would you consider revising your overall rating upward? We would be happy to clarify any remaining questions.

AuthorsOpenReview ↗

Remaining questions before author discussion closes

Dear Reviewer Jzju,

With the discussion period about to close in a day and half, could you please indicate whether the revised nonlinear evidence, humanoid experiments, and learned passive model results address your main concerns, or whether a specific issue remains unresolved?

We would greatly appreciate your updated assessment and the opportunity to answer any remaining questions before the window closes. We remain available to respond promptly.

Reviewer_RurpOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses on separating the passive dynamical process of morphological computation and the actual cost of computing the action needed to achieve a goal. In this form, the state “passive” propagation can be computed and deducted from the intervention, possibly reducing the computation cost.

Critical strengths

• Understanding the body (morphological computation) and the controller's contribution to the task-relevant information is crucial to comprehend how these systems interact and therefore provide the correct intervention or the cheapest intervention. • Formulation of the information-theoretic account to control including body contribution • Experiments on theoretical classical control examples.

Critical weaknesses

• While, in principle, this work is consistent and sound, it is not totally clear why uncoupling the passive contribution should be done; as in closed-loop control, Todorov-like but with learning, it is actually the coupling that is important, and the transition function will actually provide 0 action when the passive dynamics are expected to achieve the goal. For instance, the example of the push and coast. We could actually learn the model dynamics and alternatively use a KL regulariser on the alignment. This approach is usually taken in the expected free energy minimisation literature or model-based control as inference, where one term pushes towards the goal, and the other reduces the uncertainty of the model given a prior. • Passive dynamics may change given the action. It is not clear that the assumptions taken into account consider that the action actually could change the “passive” dynamics. In the presented experiments, the plants are ideal and do not reflect the relation between passive dynamics and actual action. For instance, a robotic arm that changes its dynamics depending on the force applied at the joint limits. • Task information: “Whether that selectivity makes control cheaper depends on the task: information about predictable motion can still be essential for correcting it.” Examples such as the mountain car reflect that the task is impacting, as the authors state, whether cheap control is useful.

Questions for authors

• Can we use this information gain split to actually help to 1) improve the controller and 2) improve the body? • Is the task always a dependent variable on the body computation?

4 comments
AuthorsOpenReview ↗

Body–controller coupling, general actuation, and task dependence

Thank you for recognizing the importance of understanding the body's and controller's contributions. Your questions sharpen the central distinction: the body selects which controller information an MC bonus compresses; the task determines whether that information can be relinquished with a physical saving. The common author comment maps the expanded evidence; below we address your specific questions.

1. Why a passive counterfactual when a coupled controller can already exploit coasting?

We agree that effective control exploits body–controller coupling. The counterfactual analyzes that coupling by asking what the same body would do from the same state and disturbance with the command withheld. The actual controlled dynamics retain the coupling. Comparing the command with this passive successor identifies which state information both carry (Section 3, Eq. 5).

This explains something that successful coasting alone does not identify: which part of feedback a regularizer encourages the controller to remove. For the coasting body in Push and Coast at 10% task degradation, MC's intervention is within 5% of Direct (137.6 vs. 131.4); Compression uses 2.7 times MC's intervention (367.4; Table 2). MC preserves alignment correction while relaxing drive feedback. Under recoverability, MC equals I(S;Y)−I(A;Y)I(S;Y)-I(A;Y) (Eq. (6)). At a fixed state ensemble, it penalizes only body-visible information; Compression penalizes the entire controller information. When the visited states change, MC can also favor more predictable passive motion.

Direct is the appropriate physical reference when intervention can be evaluated, and Corollary 32 establishes its local oracle role at matched task loss. Its presence makes the question precise: when does simplifying information align with reducing physical intervention? Section 2 places MC, empowerment, and transition KL intervention in a common geometry. Section 7 now explicitly relates this to control as inference and expected free energy. Our shared/complementary information split concerns the controller's state dependence; it is distinct from an epistemic term rewarding reduction of model uncertainty.

2. Can the action change the dynamics, particularly at joint limits?

Yes. The revision treats this directly. The controller-information split follows from independent current disturbances. The simpler MC identity additionally requires that knowing the action permits recovery of the passive successor. Proposition 11 (Appendix A.2.1) supplies the general actuation correction when that recovery fails.

We added held torque and joint stop pendulum experiments (Appendix B.5, Table 6). A stop can erase state information differently under passive and actuated motion. The resulting correction explains why full MC and the isolated BODY penalty can behave differently: full MC also reflects the changed propagation of state information. This explicitly covers the kind of action dependent mechanical constraint raised in your review.

3. Can the split improve controllers and guide body design?

The gantry crane gives a new concrete controller example: the trolley can coast, while the suspended load still needs stabilization. BODY preserves anti-sway feedback while relaxing trolley feedback, yielding 37% [36,38] less intervention than Compression; full MC saves 9% [6,11], at 10% task degradation across ten training seeds (95% bootstrap intervals; Figure 2; Appendix B.7, Table 9).

For body design, the gate identifies which state features candidate mechanics preserve, while alignment asks whether the corresponding feedback is costly and dispensable for the task. In Push and Coast, changing momentum retention changes MC's advantage over Compression from approximately 6% to 63% at the same 10% task allowance (Appendix B.1). This supplies a diagnostic for comparing candidate bodies; jointly optimizing morphology is a natural application. Learning the passive model also closely reproduces the known-model intervention results (Table 7).

4. Is the task determined by body computation?

No. The task is a specified requirement. With a common reference state ensemble, changing the task leaves the body gate unchanged but can change the physical value of relaxing feedback. We test this with four task costs on the same body (Appendix B.1). Under on-policy evaluation, the task can additionally change the visited states. Theorem 5 formalizes their interaction: predictable motion may still need correction, such as braking before a target.

We hope the explicit connection to KL control, the general actuation extension, and the controller and body comparisons resolve these concerns and support reconsideration of the overall assessment.

AuthorsOpenReview ↗

Follow up on coupling, general actuation, and task dependence

Dear Reviewer Rurp,

We would welcome your assessment of our response. We addressed all your concerns: the relationship to coupled/KL control, action-dependent dynamics and joint limits (Proposition 11), task dependence, and implications for controller and body design, supported by new nonlinear experiments. We hope these clarifications and results support a higher overall rating and would appreciate your reconsideration. We remain available for any further questions.

Reviewer_RurpOpenReview ↗

I really like the topic

Thank you very much for the effort to improve the clarity of the paper and for including the new results. I really like the topic addressed in this work related to information delegation or compression in body morphology dynamics. And I think it is good food for thought. But I re-read the paper, and I find it sometimes difficult to foresee the real impact of the work on the representation learning community. All reviewers had a similar impression, and this could probably be due to a lack of deep expertise in the authors' topic. In any case, if the rest of the reviewers think that their concerns are being totally resolved, I could reconsider a weak accept instead of a weak reject. Otherwise, I will keep my score.

AuthorsOpenReview ↗

Selective compression, information bottleneck, and physical control

Dear Reviewer Rurp,

Thank you for rereading the paper and for your openness to reconsidering the recommendation. We would like to clarify why the paper is directly a contribution to representation learning.

  • Information-theoretic regularizers connect directly to learning representations and policies. Policy compression and information bottleneck methods (Goyal et al., 2019; Leibfried & Grau Moya, 2020), alongside empowerment and skill discovery objectives (Mohamed & Rezende, 2015; Eysenbach et al., 2019), shape learned representations and behavior. Our paper asks how the body changes what an information regularizer removes from the learned controller.

  • Our central contribution is a theory of body-gated compression. Treating the action as a representation of the state, Eq. 5 separates its information into overlap with the passive successor, I(A;Y)I(A;Y), and complementary information, I(A;S∣Y)I(A;S|Y). Under recoverability and at a fixed state distribution, MC penalizes only the overlap, whereas ordinary policy compression penalizes both. The body, not the regularizer alone, selects which state information is compressed.

  • The information bottleneck connection is structural and quantitative (Remarks 1 and 3). Classical information bottleneck preserves information about a relevance variable; MC instead penalizes information shared with passive dynamics, while the task determines what must remain. In the Gaussian setting, the passive modes are precisely the bottleneck's canonical relevance directions, with their reliabilities determining the leading compression weights under weak encoding (Theorem 2). No encoder can share more than a fraction max⁡ibi\max_i b_i of its state information with the passive successor (Eq. 25). Embodiment thus explicitly shapes which state information a regularizer encourages a policy to discard.

  • The framework connects information to empowerment and to physics. State propagation and action influence occupy complementary positions in a common KL geometry (Proposition 1); empowerment describes the action channel's capacity, while intervention concerns the deployed policy (Sec. 2; Proposition 23). In the Gaussian model, intervention equals a noise-weighted quadratic control effort (Corollary 17). The mechanical energy balance separates stored energy, actuator work, and dissipation (Proposition 37), and Remark 9 shows why statistical intervention and mechanical work are distinct quantities. Theorem 5's local analysis also applies to a smooth actuator energy cost once an energy model is specified and the same regularity assumptions hold.

  • The theory makes testable predictions about physical benefit. Theorems 4 and 5 explain why information reduction alone cannot guarantee lower intervention and predict its direction and leading magnitude near a regular task optimum, including the four negative alignment cases where intervention increases (Table 5). In learned nonlinear control, original MC uses 39% [38,40] less intervention than Compression in pendulum release, and BODY uses 37% [36,38] less in the crane, at 10% task degradation across ten seeds (Tables 4 and 9).

  • Toward real robotic control with unknown dynamics, regularizers computed from a learned passive model closely reproduce the known model intervention results (Table 7). This supports a practical route: learn what the mechanics sustain from operating data, then use task–body–action alignment to guide which feedback can be relaxed within an acceptable performance loss.

  • Would this clarification of the paper's direct contribution to representation learning, together with our responses to your technical questions, support a move towards acceptance? We would welcome any specific remaining concern that we can address.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response: body-gated compression and nonlinear control

We thank all the reviewers for the thoughtful questions about nonlinear control, robustness, and the relationship between body dynamics and physical cost. The revision directly addresses these questions around one central contribution: which controller information does an MC bonus remove, and when does removing it make control physically cheaper?

Central contribution: MC as body-gated compression. Equation (5) separates controller information into the part shared with the counterfactual passive successor, I(A;Y)I(A;Y), and its complement, I(A;S∣Y)I(A;S|Y). Under recoverability, MC equals passive predictability minus the shared part (Eq. (6)); at a fixed state ensemble, an MC bonus therefore compresses the body-visible information selectively. Ordinary Compression penalizes both parts. Theorem 2 identifies the Gaussian directional weights, and Theorem 3 extends the accounting to feedback. Theorems 4–5 then separate information savings from physical savings and identify the local task–body–action alignment that connects them. Direct anchors this comparison as the physical objective and corresponding local oracle (Corollary 32).

The new evidence tests this account from controlled mechanics to neural control. Brackets denote 95% bootstrap intervals across training seeds.

  • Pendulum release: gravity supports motion toward the target, while pivot damping determines whether the body settles or needs braking. A separate alignment error still requires correction. Learned MC uses 39% [38,40] less intervention than Compression at 10% task degradation across ten seeds. Changing the pivot preserves selectivity but changes its physical benefit (Table 1; Appendix B.5, Table 4).
  • Gantry crane: trolley momentum supports travel while the suspended load needs anti-sway correction. Learned BODY relaxes travel feedback while preserving stabilization, using 37% [36,38] less intervention than Compression at 10% degradation across ten seeds; full MC saves 9% [6,11] (Figure 2; Appendix B.7, Table 9).
  • DMC Ball in Cup: the cup must launch and catch an unactuated ball. The expanded 22 seed neural policy study shows that BODY reduces body-visible information while retaining more complementary information than Compression, and identifies why intervention remains: a nearly constant post-catch holding command is physically costly but carries little state information. The comparison now includes the original MC objective alongside BODY, Compression, and Direct (Table 1; Appendix B.3, Table 3).
  • DMC humanoid balance: persistent posture still requires active correction. Ten neural controller adaptations per body, on the official and a stiffer body, extend the information–effort distinction to articulated contact dynamics. On the stiff body, information reductions of 58–65% do not translate into lower corrective effort; Direct reduces that effort by 10.6% [10.3,10.8], with no falls (Appendix B.10, Table 12).
  • Two-link arm, planar descent, and MountainCar: coupled joints require braking despite gravity assisted motion; a descending vehicle has removable effort in lateral correction; a car escaping a valley must actively build energy. These contrasting mechanics show why the task and the location of costly feedback matter alongside passive persistence (Appendices B.8, B.9, and B.6).

The revision also tests the theory's predictions and assumptions. Across nine pendulum bodies and five objectives, the local alignment prediction captures both intervention savings and the four negative alignment cases with increased intervention; quantitative agreement is strongest near the task optimum (Table 5; Figure 7). Body memory measurements show why the observation horizon matters (Figure 3). Proposition 11 extends the MC decomposition to general actuation, with held torque and joint stop experiments testing departures from recoverability (Table 6). Regularizers built from a learned passive model closely reproduce the known model intervention results in the pendulum (Table 7).

Presentation. Sections 3 now slightly revised for readability: we use consistent terms for the passive successor YY, passive future YℓY_\ell, and passive channel P0P_0, define controller information and its body-visible and complementary parts explicitly, and state the counterfactual construction of YY directly.

These additions connect the paper's exact information identities, conditional physical predictions, and empirical tests. The resulting contribution is an explanation of when mechanics can replace controller information, when that replacement saves intervention, and why the two can separate. The individual responses below address each reviewer's questions with the corresponding evidence and manuscript locations.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.