acceptodds
Under review as a conference paper at ICLR 2027

MorphAdapt: Action-Conditioned Temporal Adaptation across Unseen Robot Morphologies and Altered Dynamics

Abstract

Shared robot control faces context variation at two timescales: embodiment changes across robots, while interaction dynamics can change within deployment. MorphAdapt assigns these information roles to a static morphology pathway and a temporal adaptation pathway. A morphology-conditioned shared controller consumes context inferred by a temporal Student from recent observations and previous actions; privileged dynamics supervise training but are unavailable at deployment. We evaluate this system jointly across unseen embodiments and altered dynamics on 97 held-out UNIMAL robots spanning 87 exact-XML morphology clusters. MorphAdapt achieves 9.3%–16.4% higher mean episodic return than MetaMorph-DR in four dynamics-mutation conditions, together with lower point estimates of maximum normalized velocity drop and recovered-only recovery time. MetaMorph-DR provides a system baseline without online temporal adaptation; a matched State-only Student separately tests the value of explicit action history. Previous-action conditioning yields substantial post-perturbation point gains in an individual development realization, while its incremental effect remains heterogeneous in the matched input ablation. The results support the complete two-timescale system on this benchmark and identify reliable use of action history across training realizations as an open learning problem.

Primary area:
applications to robotics, autonomy, planning
Keywords:
robot learning, reinforcement learning, multi-embodiment control, online adaptation, legged locomotion

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_1eXvOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
2(+1)
Soundness
2
Read review
Paper summary

This paper presents MorphAdapt, a shared robot controller that combines morphology conditioning with temporal context inference to handle unseen robot embodiments and changing dynamics. A history-based Student learns from privileged Teacher supervision and adapts its inferred context during deployment without updating network parameters.

Critical strengths

Strength: Meaningful problem

The paper addresses a meaningful problem in generalist robot control: enabling a shared policy to handle both unseen robot morphologies and changes in interaction dynamics during deployment. These two sources of variation naturally coexist in practical applications, making their joint evaluation valuable.

Critical weaknesses

Weakness 1: Suspected AI-generated passages and unresolved experimental provenance substantially weaken confidence in the results.

Several passages in Appendices B and H appear AI-generated to me. In particular, unusual wording such as

“A complete Teacher completion record and historical Teacher-to-Student initialization byte receipts are not available; configured values must not be interpreted as measured Teacher completion counts. ”(Line 740-742)

reads like internal AI-assisted auditing commentary, raising concerns about whether these passages were adequately checked before submission.

More substantively, these passages explicitly acknowledge missing Teacher training completion records, initialization records, and some baseline completion/source records. Although the paper discloses AI assistance, these unresolved provenance gaps substantially undermine my confidence in the experimental reliability. The actual training budgets, checkpoint origins, and initialization procedures directly affect the interpretation and comparability of the central results.

Thus, I have assigned the highest priority to this issue, as it significantly influences my assessment of the paper's credibility.

Weakness 2: The contribution of online adaptation to the system-level improvements remains insufficiently established.

MorphAdapt achieves higher mean returns than MetaMorph-DR, but this comparison changes several factors simultaneously, including privileged supervision, Teacher-based initialization, Student distillation, and deployment-time temporal context inference. Consequently, the observed gains do not establish how much improvement comes from updating the inferred context online.

The UPDATE–FREEZE experiment in Section 6.4 provides a relevant intervention, but its results offer limited support for this central mechanism. Given the paper’s emphasis on two-timescale context and adaptation to changing dynamics, stronger evidence connecting online context updates to behavioral gains is needed to establish the distinctive value of the proposed approach.

Questions for authors

Question 1: Please clarify the meaning and origin of these unusual statements, including whether they are AI-generated text retained in the submission. What verification did the authors perform before endorsing them? More importantly, what existing evidence substantiates the completed training budgets, checkpoint provenance, and Teacher-to-Student initialization underlying the main results? (See Weakness 1)

Question 2: Can the authors isolate the contribution of deployment-time context updating within the main mutation benchmark—for example, by comparing live and frozen context using the same trained checkpoint and matched pre-mutation states and randomness? How should the weak UPDATE–FREEZE effect and negative HOLD-relative interaction be reconciled with the proposed adaptation mechanism? (See Weakness 2)

6 comments
AuthorsOpenReview ↗

Response to Reviewer 1eXv

Thank you for recognizing the importance of joint embodiment and dynamics variation. We provide specific evidence addressing experimental reliability and deployment-time context updating. The revised manuscript is being prepared for upload.

Appendix wording, AI assistance, and experimental records:

The author wrote the appendices in Chinese. AI assisted with English translation and polishing using existing training records and experimental materials as reference, as disclosed in the AI Use Statement. "Initialization byte receipts" was awkward wording for Teacher-parameter loading records during Student initialization. It was intended to distinguish configured training budgets, recorded progress, and initialization-time documentation. We take responsibility for the wording and are replacing it with standard terminology.

Documented preparation checks for the Chinese draft covered terminology, notation, compilation, references, page inspection, and preservation of table values and figure assets during editing. The submitted supplementary code archive includes implementation details; code describes the implementation, while logs and saved models document executed runs.

Additional post-submission verification establishes the following. All six Students have 99,942,400 recorded Stage-II transitions and 78,080 optimizer updates per run. Original Teacher TensorBoard records extend to 99,778,560 logged transitions. Comparing retained models identifies 131 equal corresponding non-adaptation tensors between each Student and the Teacher. We also verified 6,984 return/episode-length entries against retained evaluation outputs and reproduced 96 checked XML-cluster interval pairs within floating-point rounding.

These checks substantiate Student training counts, saved-controller parameter relationships, and reported statistics. The Teacher's complete optimizer/save history and original initialization-time loading records remain incomplete. We distinguish those historical gaps from verified saved-model relationships and statistical recomputation; neither is presented as a fresh simulator run. The prepared revision replaces the broad missing-record passage with these specific facts.

Behavioral evidence from a same-checkpoint intervention:

UPDATE and FREEZE use the same trained checkpoint, with identical prior Teacher training, initialization, and distillation. UPDATE uses current inferred context; FREEZE holds the controller-facing context at its value immediately before the second dynamics change, while history encoding and the environment continue. The outcome is closed-loop reward accumulated over the specified post-change window, not merely a latent-vector difference.

Existing results show positive mean behavioral effects in two morphology families, even though the pooled effect is small. The full confirmation sets, including negative and zero effects, give these mean UPDATE-minus-FREEZE reward differences under the RETURN schedule and HOLD-relative interactions, respectively: mvt (9 clusters), +12.515 and +5.714; vt (17), +4.073 and +4.582; floor (55), -2.989 and -6.518; all 81 clusters, +0.216 and -2.829. These are the existing morphology groups, not groups chosen by reward outcome. Positive mvt/vt contributions are offset by the larger negative contribution from floor, which accounts for 55 of 81 clusters.

The negative HOLD-relative interaction must be distinguished from the direct updating effect. With D_RETURN and D_HOLD denoting UPDATE-minus-FREEZE reward differences when motor strength returns to A or stays at B, I = D_RETURN - D_HOLD. For the same 81 clusters under equal-cluster averaging, the reported means imply D_HOLD = 0.215574 - (-2.828963) = +3.044537. Hence the negative interaction compares a smaller positive RETURN mean (+0.216) with a larger positive HOLD mean (+3.045), rather than an average penalty from updating in both schedules.

The supported result is positive mean closed-loop effects in some evaluated families, with heterogeneous effects across families and schedules. These are descriptive results from one training seed; the pooled RETURN median is zero and no uniform per-robot or cross-seed benefit is claimed. The requested same-checkpoint matched LIVE-FREEZE test on the main Mutation benchmark is not yet completed. This repeated-change intervention and the complete-system comparison address different questions; the prepared revision presents them at their respective scopes.

Thank you for your response.

Weakness 1: The authors have attempted to provide some information regarding the experiments. While this helps to some extent in establishing the authenticity of the experiments, the overly defensive tone present throughout both the main text and the appendix prevents me from changing my assessment of the work. This same defensive tone also appears in the authors' responses to Weaknesses 1 and 2. I suggest that the authors revise the language in both the appendix and the corresponding sections of the main text to ensure a higher level of academic rigor.

Furthermore, I would like to remind the authors that it is permissible to modify the main PDF during the rebuttal phase; I recommend incorporating these revisions directly into the appropriate sections of the main text.

Weakness 2: The authors state in their response:

These are descriptive results from one training seed; the pooled RETURN median is zero and no uniform per-robot or cross-seed benefit is claimed. The requested same-checkpoint matched LIVE-FREEZE test on the main Mutation benchmark is not yet completed. This repeated-change intervention and the complete-system comparison address different questions; the prepared revision presents them at their respective scopes.

This is confusing.

I would like the authors to explain the implications of the results presented in the rebuttal more directly, and to clarify how this response addresses the questions we raised.

Thank you for the clarification. We agree that our previous response emphasized qualifications before explaining the findings. We address the writing and scientific implications directly below.

Weakness 1 — Presentation and academic rigor. We will revise the main PDF, rather than rely on rebuttal text alone. Section 6.4 will state the ABA finding and its implications before experimental details; Section 7.1 will clarify what the complete-system comparisons and context intervention establish. Appendix E will retain the full results, HOLD control, and exposure definitions while removing repetitive, defensive, and internal-audit-style prose. We will similarly edit the relevant implementation and baseline descriptions. We will state limitations where they affect interpretation, not as substitutes for explaining the results.

Weakness 2 — What the interventions show. Our evidence shows that changing the temporal context supplied to the controller can improve locomotion, and that continued updating helps in some tested cases. It does not yet identify a positive average benefit of online updating across the main benchmark.

ABA: Motor strength changes from 1.0 to 0.8 at index 250 and returns to 1.0 at 550. In [550,800), UPDATE uses the newly inferred context; FREEZE holds the pre-restoration context. The matched HOLD branch retains motor strength 0.8. Across 81 confirmation XML clusters, the mean UPDATE–FREEZE reward difference on RETURN is +0.216 (median 0). Family means are +12.515 (mvt), +4.073 (vt), and −2.989 (floor): the response varies across robot families. The RETURN-minus-HOLD interaction is −2.829; equivalently, the mean updating advantage is +3.045 in HOLD versus +0.216 in RETURN. Thus updating can help in some conditions, but motor-strength restoration does not produce an additional average updating benefit in this test.

GearPatch pilot: We further tested an exploratory intervention that divides the history encoder's gear-feature input-weight columns by 300, without changing actuator gear, the action decoder, or performing policy optimization. The pilot uses four previously evaluated morphologies, evaluation seed 1409, and a 500-step horizon. We compare Original, GearPatch-Live, and GearPatch-Frozen; the latter two share the patch and differ only in whether the decoder receives updated context from decision 250 onward. The full pilot covers 144 episodes across State-only and State+Action, and all 48 patched Live/Frozen pairs have matching pre-freeze traces.

For State+Action, the Mutation Return contrasts averaged over all four morphologies are:

  • Seed 1409: Live−Original +712.44; Live−Frozen −50.35.
  • Seed 1410: +43.42; −51.41.
  • Seed 1411: +402.07; +7.09.

The patched policy therefore improves mean Mutation Return over Original in all three Student runs. The gains include longer survival: +628.00 of the +712.44 gain for seed 1409 comes from additional surviving steps. We also observe a positive within-policy updating case for a seed-1411 mvt robot: Live Return 2757.59 versus Frozen 2542.32, both reaching 500 steps. The original latent is zero in these pilot trajectories, whereas the patched context is nonzero. In same-state probes, reversing earlier history changes the patched latent in 204/205 snapshots; substituting patched for original context changes executed actions in 202/205 snapshots. These are sensitivity probes, not independent robot trials.

Crucially, the Frozen arm retains substantial gains because it retains the patch and the pre-change latent. Thus the large Live−Original gains demonstrate behavioral effects of the context configuration, while Live−Frozen measures the additional value of continued updating and yields mixed results (negative in two of three State+Action seeds). The pilot and ABA also use different interventions and protocols; they are not conflicting estimates of one effect.

Implications and revision. The experiments show meaningful effects of the context pathway and local benefits from updating, but do not establish that continuous latent updates account for the main system-level Return advantage over MetaMorph-DR or R-MoMo. The requested main-protocol LIVE–FREEZE study remains paused because its LIVE replay does not reproduce historical Return; no formal effect is claimed from it. The final revised manuscript will include the GearPatch protocol, all per-seed and per-morphology results, and paired Live/Frozen controls as supplementary material, with the key interpretation incorporated into the main Results and Discussion.

Reviewer_Au5WOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes MorphAdapt, a shared-control system for heterogeneous morphologies that separates static nominal embodiment context from a time-varying, action-conditioned interaction context. The controller is a morphology-conditioned shared Transformer over padded limb tokens; a causal history encoder infers a 16-D temporal latent from a fixed window of observations and previous canonical actions. A privileged Teacher, trained with PPO, encodes runtime dynamics parameters; a Student is then trained by interactive distillation with latent MSE and action-mean MSE, while the shared controller remains frozen and only the Student policy history encoder is updated. At deployment, the Student uses nominal morphology, current observations, masks, and recent history, without privileged dynamics, Teacher targets, or online gradient updates.

The empirical validation uses UNIMAL morphologies in MuJoCo, with 100 training robot instances and Strict-OOD97 held-out walkers in 87 exact-XML clusters. The main benchmark crosses Fixed and Mutation dynamics with nominal, ID, OOD-mild, and OOD-strong reset conditions. The paper reports 9.3%–16.4% higher mean return under mutation than MetaMorph-DR and lower maximum normalized post-mutation velocity drop than compared systems. It also provides a matched State+Action vs State-only input ablation, horizon sensitivity, repeated-change diagnostics, structural-distance analysis, and fallen-state robustness checks. The paper is candid that the action-input ablation does not show a consistent benefit and that system-level comparisons include training-pipeline differences.

Critical strengths

1.The evaluation protocol is substantially more careful than a typical system-level benchmark.

The use of exact-XML-held-out morphology clusters, three training seeds for the main comparison, cluster bootstrap intervals, matched SO/SA input ablation, and explicit diagnostics for mutation recovery and repeated context updates gives the paper credible internal validation. This is consequential because the paper’s main claim is system-level robustness under unseen morphology and altered dynamics; without this protocol, the contribution would be much weaker.

2.The deployment formulation is practically meaningful and the reported mutation robustness is nontrivial.

The Student requires neither privileged dynamics nor Teacher targets nor online parameter updates, and yet reports higher mean return under all four mutation conditions and lower maximum normalized post-mutation velocity drop than the compared systems. This is consequential because it supports the paper’s central practical claim that online context inference can improve changing-dynamics control on held-out embodiments without deployment-time adaptation of network parameters.

Critical weaknesses

1.The methodological novelty is limited, and the necessity of the paper’s central action-conditioned temporal component is not convincingly established.

Morphology-conditioned shared Transformers, history-based adaptation, privileged teacher-student distillation, and action-conditioned history are all closely related to prior work; Eq. (4) explicitly follows privileged locomotion imitation. The paper’s Eq. (1) is an information-role distinction rather than a learned-independence guarantee. More importantly, the matched SA–SO ablation reports return-difference intervals that all include zero, and the MND direction is mixed under mutation. This is consequential because the paper’s claimed contribution is not merely integration but an action-conditioned temporal context for changing dynamics; if explicit action history does not yield consistent gains, the core technical novelty and necessity need stronger justification.

2.The baseline suite is dominated by older methods and does not fully reproduce all structures/configurations of the newer hypernetwork-style comparison, while baseline training protocols and fairness are insufficiently disclosed.

MetaMorph and ModuMorph are older shared-morphology baselines; HyperDistill is included but only for the four mutation cells, with an adapted MetaMorph-DR Teacher, a single distillation seed, and missing fixed-dynamics cells. ModuMorph is a single-seed released-code reproduction. The paper does not clearly report, for each baseline, the number of training morphologies, total training budget, checkpoint/stage selection, tuning budget, whether the same reset dynamics randomization was used, and matched evaluation seeds. MetaMorph-DR is also not seed-matched with SA. This is consequential because the main comparisons are system-level and include training-pipeline differences; without these controls, the reported gains cannot be cleanly attributed to the proposed temporal adaptation rather than to baseline training budgets, checkpoint selection, or incomplete reproduction.

Questions for authors

1.Baseline protocol table (clarification).

Could the authors provide a table listing, for MetaMorph-DR, ModuMorph, and HyperDistill: number of training morphologies, total environment transitions, training seeds, whether privileged information was used, whether online adaptation was used, checkpoint/stage selection criterion, evaluation seeds, and tuning budget? Were all baselines trained with the same reset dynamics randomization ranges as MorphAdapt? This materially affects whether the system-level gains are attributable to the proposed method.

2.HyperDistill reproduction completeness (clarification).

The paper states that HyperDistill uses an adapted MetaMorph-DR Teacher rather than the original paper’s universal Transformer-ModuMorph Teacher, a single distillation seed, and only mutation cells. Could the authors clarify whether the full HyperDistill architecture and training structure were reproduced, and why fixed-dynamics cells are missing? If fixed cells are unavailable, how should readers interpret the claim that MorphAdapt exceeds HyperDistill in all four mutation conditions?

3.Evidence for action-conditioned temporal context (clarification).

Given that all SA–SO return-difference intervals include zero and the MND direction is mixed under mutation, what evidence supports explicit action history as a necessary component rather than state history alone? Could the authors provide a seed-matched SA–SO comparison with additional seeds, or a mechanism analysis showing when action history changes the inferred latent in a behaviorally relevant way?

4.Novelty relative to prior methods (clarification).

Beyond integrating static morphology tokens, history-based adaptation, and privileged distillation, what is the precise conceptual or technical difference from MetaMorph/ModuMorph, RMA-style privileged distillation, and Lee et al.-style temporal Student imitation? Which ablation isolates this difference from the shared controller and training pipeline?

1 comment
AuthorsOpenReview ↗

Response to Reviewer Au5W

Thank you for the detailed review. The revision is being prepared for upload; the original submission remains on OpenReview.

Baseline protocols and budgets:

Appendices B/C/H and the submitted code archive contain several requested details. Below we consolidate them with verified records; counts are per run or stage.

MorphAdapt State-only (SO) and State+Action (SA) each train on 100 robots with seeds 1409/1410/1411: 99,942,400 transitions and 78,080 updates per Student. Their shared privileged Teacher separately logs 99,778,560 transitions (configuration seed 1409). MetaMorph-DR uses the same 100 robots, seeds 1412/1413/1415, and 99,942,400 transitions per run (78,080/78,028/77,999 updates). The released-code ModuMorph reproduction uses 100 robots, seed 1409, and 99,942,400 transitions (100,885 reconstructed updates). Adapted HyperDistill uses a DR Teacher trained on 100 robots (seed 1415), plus 8,000,000 expert transitions from 1,000 static variants of 100 parents; seed-1409 distillation runs 150 epochs with 234,450 updates. Evaluated checkpoints are terminal Student/DR at 1,220 iterations, ModuMorph at index 1219, and HyperDistill at epoch 150. Repeated sample uses are not new environment transitions.

SA/SO and DR share training robot IDs and episode-constant reset multipliers: motor/mass-inertia [0.8,1.2] and friction [0.7,1.3]. ModuMorph did not use matched reset randomization; HyperDistill uses a DR Teacher and separate morphology-variant collection. MorphAdapt uses privileged dynamics for Teacher training and Teacher latent/action targets for Student training, but only observable history at deployment. DR has no temporal adaptation module; HyperDistill generates policy weights once per episode. ModuMorph's actor/critic privileged inputs are not fully established. All deployment weights are fixed.

All baseline comparisons use evaluation seed 1409 on the common held-out protocol. ModuMorph covers eight conditions and HyperDistill four Mutation conditions plus Fixed-Nominal. Complete tuning and checkpoint-selection histories are unavailable. SA-SO is seed-matched; SA-DR has unpaired seeds, different supervision, and additional Teacher cost.

HyperDistill implementation and Fixed coverage:

Appendix H already identifies our MetaMorph-DR Teacher adaptation. We retain morphology-conditioned hypernetwork-generated MLPs, observation conversion, task-balanced distillation, and context-embedding dropout, but do not claim to reproduce the original training pipeline unchanged. Original Table 8 includes Fixed-Nominal (mean return 1025.3, 97 episodes). The main figure shows four Mutation conditions; the other three Fixed results are unavailable. The comparison concerns the evaluated adapted checkpoint, not HyperDistill generally. The prepared revision clarifies this.

Action history:

Section 6.3 and original Table 3 already report three matched SA/SO training seeds. The variants share the Teacher, temporal backbone, H=10, latent width, optimizer, budget, and evaluation protocol. SA adds previous actions and 768 trainable encoder weights. All eight Return-difference intervals and four Mutation MND-difference intervals include zero; directions vary across conditions. This tests explicit actions, not temporal inference itself: SO also infers context. The evidence establishes neither necessity nor equivalence. The prepared revision adds existing per-seed effects and denominators; no new seeds or action-specific intervention are claimed.

Scientific question and evidence:

MorphAdapt studies joint generalization to unseen bodies and changing deployment dynamics. It supplies nominal morphology explicitly and infers episode-local context from history, fused with limb representations at the decoder. Training uses episode-constant dynamics; evaluation introduces within-episode changes on held-out bodies. Equation (1) defines information roles; Equation (4) follows privileged temporal imitation. SA-DR compares complete systems, SA-SO tests action input, and UPDATE-FREEZE probes context use within a fixed checkpoint on a separate dynamics schedule. These comparisons answer different questions; they do not fully isolate the two-pathway architecture from the entire training pipeline. SO itself remains a history-conditioned shared controller, so action-input necessity is narrower than the joint-generalization question.

Recent baseline:

We completed a one-seed reproduction and native evaluation of Shared Modular Recurrence in Contextual MDPs for Universal Morphology Control, cited in Related Work. We are preparing evaluation under MorphAdapt's held-out morphology and mutation protocol. Native results are not matched numerical comparisons; we will report these only after verification.

Reviewer_3AKiOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a method to learn a unified controller that generalizes across diverse morphological variations for locomotion tasks. The core methodological contribution is a teacher-student framework that accounts for dynamics (earlier work, like MetaMorph, only conditions on morphology and uses domain randomization to handle dynamic variations). The results show moderate improvement over MetaMorph, with greater gains in OOD scenarios.

Critical strengths
  • Incorporating dynamics into learning can provide better policies than domain randomization, which can get stuck at suboptimal solutions that work for all cases, instead of adapting to dynamics. That is, I agree with the motivation, and the approach is reasonable.
Critical weaknesses

My major concerns are: i) whether the marginal improvement provides any real benefits, and ii) novelty.

  • The paper uses the setup from MetaMorph and is mainly centered around simulation and non-realistic morphologies. The practical benefits of the approach are unclear: whether the small to moderate improvement over MetaMorph translates to any benefit in the real world.
  • The technical contribution of using a teacher-student approach to adapt to unknown environment factors is not novel in locomotion - Rapid Motor Adaptation (https://arxiv.org/abs/2107.04034), RSS 2021. Combining it with a multi-task learning setting across diverse morphologies is very incremental, given that the multi-morphological setting (MetaMorph) already exists.

Minor

  • some details can ideally be moved to the supplementary material (such as MLP layers, the exact seed values) L174.
  • the writing style ("it is not ...") (L158, L254, etc.) is confusing at times. The writing is very hard to follow: Unimportant details are embedded in the sentences, making the content hard to follow.
  • SO and SA short forms are used without describing what they are. (state only and state+action)
  • An overview diagram for the method will make understanding the two branches easier.
Questions for authors
  • see above.
1 comment
AuthorsOpenReview ↗

Response to Reviewer 3AKi

Thank you for recognizing the motivation for conditioning shared control on changing dynamics. We would like to clarify the scientific question underlying MorphAdapt, the evidence addressing that question, and the presentation changes being prepared.

  1. The central question is joint generalization across embodiment and changing dynamics

MorphAdapt asks: can a single controller trained across heterogeneous robots maintain effective locomotion on unseen bodies when their dynamics change during deployment, without robot-specific retraining or online parameter updates?

This question is not resolved by establishing morphology generalization and dynamics adaptation separately. Embodiment affects the relationship between control inputs and state transitions. Consequently, temporal inference learned across training robots must also operate under embodiment variation, while the shared policy must respond to changes occurring within an episode. The existence of a morphology-sharing method and an adaptation method does not establish their performance under these joint changes.

MorphAdapt studies this problem by distinguishing two sources of control information: known nominal embodiment and interaction-dependent context. Nominal morphology and current limb state remain explicitly available to the shared Transformer, while a temporal Student infers a 16-dimensional context from an H=10H=10 observation–previous-action window. The context is fused with the limb representations at the decoder. Privileged supervision trains this inference across bodies; deployment uses the resulting shared controller with fixed weights.

The experimental design directly examines the associated generalization question. Training uses 100 robot instances with randomized dynamics held constant within each episode. Evaluation covers 97 held-out instances in 87 exact-XML clusters and introduces persistent within-episode dynamics changes. Both the robot instances and the temporal dynamics regime therefore differ from training. Our contribution is a method and empirical investigation of shared control under this joint variation, with separate comparisons for complete-system performance, action-history inputs, and context updating.

  1. The measured benefits concern both task performance and perturbation response

Across the four mutation conditions, MorphAdapt achieves 9.3%–16.4% higher mean episodic return than MetaMorph-DR. Maximum normalized velocity-drop (MND) point estimates are also lower in all four conditions. For nominal and OOD-mild resets, MND is 1.157 versus 1.575 and 1.373 versus 1.895, respectively. These comparisons measure both accumulated task reward and the largest relative loss of forward locomotion following a dynamics change.

The 95% XML-cluster bootstrap intervals exclude zero for the Mutation OOD-strong return difference and the nominal/OOD-mild MND differences; the remaining mutation intervals include zero. These intervals quantify morphology uncertainty conditional on the evaluated models.

UNIMAL enables controlled comparisons across heterogeneous limb configurations under motor-strength, friction, and mass/inertia changes. Its role is to expose cross-embodiment behavior under a common protocol, rather than to reproduce a particular commercial robot. The practical objective is reuse of a trained controller across bodies and changing operating conditions. The current experiments establish performance in this simulation setting; physical-robot transfer requires separate validation.

  1. Presentation and implementation details

The original submission includes the method overview in Figure 1 and the SO/SA definitions in Section 4.2, with further implementation details in the appendices and supplementary code archive.

To make the main presentation easier to follow, the revision being prepared for upload brings the definitions forward, presents the two information pathways and their decoder fusion more explicitly, and distinguishes Teacher training, Student distillation, and deployment in Figures 1 and 5. It also consolidates secondary architecture details and exact seed identifiers in implementation tables and replaces the cited negative constructions with direct method and experiment descriptions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.