PaperlayerView exact revision on arXiv

Source-linked AI summary

When Instructions Retrieve Trajectories: Diagnosing and Mitigating Generalization Failures in VLA Models

Hung-Jen Chen, Yu-Hsun Hou, Yan-Hong Chen, Yan-Fu Chen, Binghua Cai, Min Sun, Chun-Yi Lee

arXiv:2609.39971v1cs.ROcs.LG

TL;DR

VLA models can succeed on nuisance changes yet fail when counterfactual changes require a different action, with familiar-behavior failures despite responsiveness to language and vision. The paper analyzes instruction-action binding and introduces ECT, which supplies and jointly trains same-instruction alternatives. CALVIN and UR5e results show improved completion and unseen-position success in their evaluated settings.

  • Problem

    VLA policies can perform well on nuisance changes yet fail when counterfactual changes require a different action, raising whether success under changed input demonstrates action selection appropriate to the instruction and scene.

    Source: §1 · p. 1

  • Method

    ECT supplies valid demonstrations in distinguishable scenes requiring different actions under the same instruction and trains each demonstration with its counterpart in the same update.

    Source: §1 · p. 2

  • Results

    58% to 76% five-task completion on CALVIN, while UR5e unseen-position success rises from 3/40 to 35/40 with 400 demonstrations per main model.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The experiments examine constructing missing action alternatives, exploiting existing counterparts, and collecting scene-dependent demonstrations across LIBERO, CALVIN, and UR5e.

    Source: §1 · p. 2, §6 · p. 10

    ECT improves distinctions exposed by added alternatives but does not create distinctions that remain absent from the training data.

    Source: §5.2 · p. 8

Abstract

from arXiv · show

Vision-language-action (VLA) models can exceed 90% success on in-distribution tasks and withstand nuisance changes that preserve the required action, yet fail under counterfactual changes that demand a different action. Aggregate robustness scores can therefore conceal a more specific failure, in which a policy responds to both language and vision yet does not combine them to select the action the task requires. We call this failure instruction-action binding. Instructions cue familiar trajectory families, and visual feedback adjusts their execution. Behavioral analyses of fine-tuned $\pi_{0.5}$ and GR00T-N1.7 policies reveal that failed rollouts often retain the source behavior or switch to another demonstrated task. These switches show that language is not simply ignored. Readouts and interventions connect these choices to task-conditioned internal states. Our analysis of the imitation objective shows how narrow conditional action support can leave grounded and instruction-keyed solutions indistinguishable on the demonstrations. This motivates Equivariant Counterfactual Training (ECT), which acts at two levels. ECT data supply valid demonstrations in which the same instruction requires different actions in distinguishable scenes, while the ECT loss trains each demonstration with its counterpart in the same update. In a controlled LIBERO-PRO comparison, full ECT raises $\pi_{0.5}$'s mean position-swap success from 36% to 59%. On CALVIN, where counterparts already occur in the original data, the ECT loss improves five-task completion without new demonstrations. On a real UR5e under a fixed demonstration budget, full ECT raises unseen-position success from 8% to 88%.

1 INTRODUCTION

VLA policies can remain responsive to language and vision while selecting a familiar behavior that no longer satisfies the task. The paper diagnoses this as instruction-action binding and motivates ECT with evidence that narrow conditional action support permits instruction-keyed solutions.

1 INTRODUCTION

  • Nuisance changes preserve the action required for success, whereas counterfactual changes require a different action; the latter expose poorer performance across four LIBERO-PRO suites.The paper reports this pattern for both π0.5 and GR00T-N1.7.Source: §1 · p. 1
  • 69% of targeted π0.5 probes followed a familiar behavior, while only 4% collapsed without a coherent trajectory.The paper reports that these failures can retain the source task or switch to another demonstrated task.Source: §1 · pp. 1–2
  • Narrow trajectory families in task-specific demonstrations can let an instruction-keyed solution fit supervision without learning scene-dependent action changes.Across a fixed-step sweep from 25% to 100% of the original data, every Swap and Task cell remains at least 34 points below in-distribution success.Source: §1 · p. 2
  • ECT supplies same-instruction demonstrations in distinguishable scenes requiring different actions, then trains each demonstration with its counterpart in the same update.The paper evaluates constructed LIBERO counterparts, naturally occurring CALVIN alternatives, and physical UR5e demonstrations.Source: §1 · p. 2
  • On CALVIN, five-task completion rises from 58% to 76%; on UR5e, unseen-position success rises from 3/40 to 35/40 with 400 demonstrations per main model.These settings use naturally occurring alternatives and physical demonstrations, respectively.Source: §1 · p. 2

2 A FRAMEWORK FOR INSTRUCTION-ACTION BINDING

The framework distinguishes grounded action selection from instruction-keyed trajectory lookup. It relates their indistinguishability on narrow-support demonstrations to the paper’s account of instruction-action binding and its testable implications.

2 A FRAMEWORK FOR INSTRUCTION-ACTION BINDING

  • In LIBERO fine-tuning demonstrations, the same instruction usually corresponds to a narrow trajectory family.The instruction usually identifies the task, and outside Goal the scene layout usually does too.Source: §2 · p. 3
  • A grounded policy resolves an instruction’s referent in the scene, while an idealized lookup selects a demonstrated trajectory family from the instruction.The framework notes that actual task-conditioned states can depend on both vision and language.Source: §2 · p. 3

2.1 CONDITIONAL ACTION SUPPORT

  • With narrow within-task trajectory families, an instruction-keyed action solution can fit demonstrations without resolving scene-dependent action changes.The paper’s flow-matching analysis states that the conditional target changes little when the scene and instruction are replaced by the instruction key.Source: §2.1 · p. 3
  • ECT supplies valid alternatives under one instruction, pairing distinguishable observations that require different actions to supervise scene-dependent action choice.The alternatives provide selection supervision that an instruction-only model cannot obtain from the instruction alone.Source: §2.1 · p. 3
  • The indicators Δk and Δa separate changes in the selected instruction key from changes in the action required for success.A Swap preserves the key but changes the required action; a Task change can retain or switch the key.Source: §2.1 · pp. 3–4
  • The table’s Semantic case represents paraphrases that preserve the instruction key, while a Task change can retain or switch it.The framework cautions that a key change is not itself evidence of correct task following.Source: Table 1 · p. 4, §2.1 · p. 4

2.2 TESTABLE IMPLICATIONS

  • The framework tests whether nuisance performance stays high while counterfactual changes expose larger gaps and failures resemble familiar behaviors.It also tests retained local tracking, task-conditioned state influence on action selection, and whether alternatives help where repetitions do not.Source: §2.2 · p. 4

3 COUNTERFACTUAL FAILURES FOLLOW FAMILIAR BEHAVIORS

Official-checkpoint evaluations show lower success on counterfactual than nuisance cells for both tested architectures. Behavioral probes further show familiar-task retrieval and retained visual tracking of targets despite incorrect task-level choices.

3 COUNTERFACTUAL FAILURES FOLLOW FAMILIAR BEHAVIORS

  • On both architectures, every nuisance cell exceeds every counterfactual cell within each suite, and Swap and Task are at least 42 points below ID.The evaluation covers LIBERO-PRO Spatial, Object, Goal, and Long (LIBERO-10).Source: §3 · p. 4

3.1 SELECTING A FAMILIAR BEHAVIOR, THEN TRACKING ITS TARGET

  • Of 381 π0.5 rollouts, 69% follow a familiar behavior and 27% follow or approximately follow the changed requirement.After excluding two rules where familiar behavior can remain valid, retrieval-like behavior is 87% across 291 probes; binary-label agreement is κ = 0.78.Source: §3.1 · p. 5
  • On the same 381 probes, GR00T-N1.7 is 73% retrieval-like and 14% correct or near-correct; π0.5 often switches tasks on Goal and retains source behavior on Long.A nearest-trajectory check agrees with human labels on 88% of usable Goal retrieval cases.Source: §3.1 · p. 5
  • In all nine combined swapped-and-shifted Object probes, π0.5 tracks the displaced object in the source-bound slot rather than the instructed object.The policy adjusts execution with visual feedback while retaining the wrong task-level choice.Source: §3.1 · p. 5

4 TASK-CONDITIONED STATES INFLUENCE ACTION SELECTION

Readouts and interventions test whether task-conditioned internal states relate to the familiar behaviors selected under changed instructions. Across π0.5 and GR00T-N1.7, interventions redirect motion toward a target task, though they do not make the injected target task succeed.

4.1 READOUT AND INTERVENTION ON TASK-CONDITIONED STATES

  • For fixed-observation Goal rollouts, the executed task ranks first in 34/45 prefix-KV readouts, compared with 22/45 for text-only TF-IDF.Holding the observation fixed makes the readout reflect instruction-dependent task selection beyond surface-word similarity.Source: §4.1 · p. 5
  • Changing task-conditioned states redirects action selection toward the target task in both π0.5 and GR00T-N1.7, but neither architecture completes the injected task.For π0.5, target-directed motion occurred in 109/200 rollouts versus 0/200 controls; for GR00T-N1.7, it occurred in 121/200 versus none of the controls.Source: §4.1 · p. 5
  • The closed-loop intervention replaces the prefix-KV cache at every policy call with one computed from the target task’s initial observation and instruction.Matched controls retain the source cache, while source scene, instruction, and policy remain fixed.Source: §4.1 · p. 5
  • In a single-call intervention, patched actions align more with the patched task’s expert trajectory than with other tasks, with cosine 0.44 versus 0.18.The intervention holds the observation fixed while substituting a candidate instruction’s final-layer cache.Source: §4.1 · p. 5

5 EQUIVARIANT COUNTERFACTUAL TRAINING

ECT supplies valid same-instruction demonstrations requiring different actions in distinguishable scenes, then trains each counterpart pair together. The section reports gains across LIBERO-PRO, CALVIN, and a fixed-budget UR5e study, while showing that improvements depend on which distinctions the alternatives expose.

5.1 METHOD

  • A useful counterfactual holds the instruction fixed, changes the required action, and makes the scene reveal which action is appropriate.Front-back mirroring alone left landmarks in similar image regions and reduced diagnostic Object ID success from 94% to 24%; added shifts or other mirrors can separate the relevant landmarks.Source: §5.1 · pp. 6–7
  • ECT loss trains matched same-instruction counterparts by summing their standard imitation losses in the same parameter update.It adds no auxiliary objective; the π0.5 implementation also shares flow-matching noise and flow time across pair members.Source: §5.1 · p. 7

5.2 CROSS-ARCHITECTURE RESULTS

  • Beyond directly constructed spatial changes, ECT improves Goal Task by 15 points for Frozen-LM π0.5 and Object Task by 17 points for Full FT.Across 381 probes, changed-requirement behavior rises from 27% to 48%, while retrieval-like behavior falls from 69% to 47%; this comparison also changes the fine-tuning recipe.Source: §5.2 · p. 7
  • Mean π0.5 Swap success rises from 36% to 59%, and mean Task success from 21% to 28%, in the controlled ECT comparison.The Frozen-LM models use matched steps and batch size; the table also includes comparisons to official Full-FT and GR00T checkpoints under different conditions.Source: §5.2 · p. 7, Table 2 · p. 8
  • ECT’s adaptation is limited to distinctions exposed by its alternatives: both models still choose the bowl after an object-name change when Spatial training always used bowls.For new Spatial destinations, erroneous completion of the original placement falls from 29/30 to 3/30.Source: §5.2 · p. 8

5.3 BEYOND LIBERO, AND BEYOND MORE DATA

  • With 400 demonstrations per main model, unseen-position UR5e success is 3/40 for Standard, 30/40 for ECT data, and 35/40 for the paired method.The unseen positions occur in none of the training layouts.Source: §5.3 · p. 8
  • On CALVIN ABC→D, ECT loss alone raises five-task completion from 58% to 76% using counterparts already present in the demonstrations.No additional demonstrations are constructed for this loss-only comparison.Source: §5.3 · p. 8
  • Increasing original LIBERO demonstrations from 25% to 100% still leaves every Swap and Task cell at least 34 points below in-distribution success.This fixed-30k-step comparison distinguishes scene-dependent alternatives from demonstration count alone.Source: §5.3 · p. 8

6 RELATED WORK AND CONCLUSION

The paper positions instruction-action binding alongside work on VLA generalization, shortcut learning, counterfactual supervision, and robot demonstration construction. It concludes that ECT targets the supervision gap with distinguishable scene-dependent alternatives and paired training, and that counterfactual evaluation tests task-level grounding.

RELATED WORK

  • Unlike Counterfactual Action Guidance, which strengthens language conditioning at inference, ECT supplies scene-dependent action alternatives and trains counterparts together.The paper’s probes and interventions show that instructions can select a wrong familiar behavior while local visual tracking remains.Source: §6 · p. 9
  • ECT differs from CAST by fixing the instruction while varying the scene and required action, and it keeps the policy architecture unchanged.It shares geometric demonstration-construction mechanisms with RoCoDA and MimicGen; successful replay and observable scene differences support action-valid, distinguishable alternatives.Source: §6 · p. 9

CONCLUSION

  • The conclusion identifies narrow conditional action support as permitting instruction-keyed solutions and presents ECT data plus paired ECT loss as responses to that supervision gap.Across LIBERO, CALVIN, and UR5e, the paper describes constructing missing alternatives, exploiting existing counterparts, and collecting scene-dependent demonstrations.Source: §6 · p. 10

NOTATION TABLE

The supplied passage identifies a notation table but does not provide its entries.

NOTATION TABLE

  • Table 5 is described as listing frequently used notation in the main paper and appendix.Source: Table 5 · p. 14

SUPPLEMENTARY OVERVIEW

The appendix outlines the evaluation setup and supporting analyses, including two VLA families, four LIBERO suites, and terminal-state success scoring.

SUPPLEMENTARY OVERVIEW

  • The appendix covers evaluation corrections, objective analysis, behavior and mechanism studies, ECT construction and results, statistical checks, related accounts, and limitations.Its sections also include the physical UR5e and CALVIN studies.Source: SUPPLEMENTARY OVERVIEW · p. 14

A EXPERIMENTAL DETAILS

  • The evaluation studies π0.5 and GR00T-N1.7 across Spatial, Object, Goal, and Long, with LIBERO-PRO perturbation cells including Semantic, object perturbation, Swap, and Task.Semantic and object perturbations usually preserve the required trajectory; Swap and Task require recomputing actions from the current scene and instruction.Source: A EXPERIMENTAL DETAILS · pp. 14–15
  • LIBERO success is scored by the terminal-state predicate, with 500 trials per evaluation cell across 10 tasks per suite.Human rollout labels and mechanistic diagnostics are separate analyses from simulator success.Source: A EXPERIMENTAL DETAILS · p. 15, Table 6 · p. 15

B LIBERO-PRO EVALUATION PROTOCOL

This material clarifies how LIBERO-PRO instructions are delivered and develops the objective-level and behavioral analyses of instruction-action binding.

B LIBERO-PRO EVALUATION PROTOCOL

  • The LIBERO-PRO protocol reads Semantic and Task instructions from the regenerated BDDL language block, correcting evaluation code that otherwise supplied the original filename-derived instruction.The other listed cells retain filename-derived wording; simulator goals and success predicates are unchanged.Source: B LIBERO-PRO EVALUATION PROTOCOL · p. 15

C OBJECTIVE-LEVEL INTERPRETATION

  • When instruction identity nearly determines the demonstrated trajectory, the flow-matching objective can fit the demonstrations without requiring counterfactual scene-conditioned recomposition.The conditional-optimal velocity is an expectation over the demonstrated action and noise given the interpolated action, time, scene, and instruction; under the stated condition it is approximately instruction-keyed.Source: C OBJECTIVE-LEVEL INTERPRETATION · pp. 15–16

D.1 ANNOTATION PROTOCOL

  • Human annotators compare perturbed rollouts with source, requested, and other demonstrated behaviors, grouping labels into retrieval-like behavior, following, and collapse.A primary annotator labeled all 381 rollouts; a second independently labeled 117 as a reliability check, with agreement varying by label granularity.Source: §D.1 · pp. 16–17, Table 8 · p. 16, Table 9 · p. 17

D.2 AGGREGATE AND RULE-LEVEL OUTCOMES

  • For both official policies, whose action pathways share no component, following a familiar trajectory family is dominant; GR00T-N1.7 collapses more often than π0.5, mostly on Goal.The same 381 probes were labeled with the same rubric for official GR00T-N1.7 and π0.5 trained with ECT.Source: §D.1 · p. 17
  • On 381 π0.5 probes, 69.3% show source-task, other-task, or hybrid retrieval, while 26.5% show correct or near-correct following.The table aggregates human-labeled outcomes by rule and suite.Source: §D.2 · p. 17, Table 10 · p. 17

D.3 SELECTED BEHAVIORAL CASE STUDIES

The case studies show familiar-task retrieval in perturbed rollouts and report how ECT changes behavior on the same probes.

D.3 SELECTED BEHAVIORAL CASE STUDIES

  • Under Goal Rg,4, 21 of 30 rollouts execute another familiar demonstrated task, while all 30 Spatial Rs,3 rollouts retain the source placement trajectory.The combined Goal perturbation changes several task fields; Rs,3 changes the destination.Source: §D.3 · pp. 17–18
  • In an unchanged-instruction object stress test, all nine rollouts select the object in the source-bound pickup slot and track it smoothly after a 10 cm shift.The displacement-following behavior distinguishes the result from fixed-coordinate replay.Source: §D.3 · p. 18, Table 12 · p. 18

D.4 THE SAME PROBES AFTER ECT

  • ECT-trained π0.5 carries out the changed task in 47.5% of the 381 probes, compared with 26.5% for official π0.5, with gains on every suite.Both policies complete unperturbed source tasks at 95% or more; under a new Spatial destination, source-task completion is 3 of 30 for ECT versus 29 of 30 for official π0.5.Source: §D.4 · p. 18

E MECHANISTIC DIAGNOSTIC DETAILS

E.1 PREFIX-KV RETRIEVAL PROTOCOL AND ROBUSTNESS

The appendix tests how task-conditioned prefix representations relate to action decoding and details ECT's data construction, intervention, and validity criteria.

E.2 PREFIX-KV PATCHING INTERVENTION

  • Restoring the source-task prefix moves the chunk-level decode closer to the source expert in 87% of rollouts, while all-layer patching changes the chunk more but worsens first-step distance.A larger intervention therefore need not improve both measures.Source: §E.2 · pp. 20–21, Table 16 · p. 21

E.1 PREFIX-KV RETRIEVAL PROTOCOL AND ROBUSTNESS

  • The executed task is the nearest prefix-KV candidate in 34 of 45 rollouts, versus 7 of 45 in a mixed-scene control; alternative aggregations preserve the retrieval signal.Each query compares candidate task instructions under the same source scene; each layer reaches at least 60% rank-1 recovery.Source: §E.1 · p. 19, Figure 9 · p. 22, Figure 10 · p. 22

E.2 PREFIX-KV PATCHING INTERVENTION

  • Persistent target-state injection redirects motion toward the target object in over half of rollouts and in no controls, but no injected rollout completes the target task.The injected representation is computed from the target task’s initial observation rather than updated from the live scene.Source: §E.2 · pp. 21–22

F ECT DATA CONSTRUCTION

  • ECT constructs scene-transformed counterparts, using mirrors or shifts and replaying transformed paths in the changed scene to obtain action labels.The transformation operators are illustrated with the robot unchanged.Source: §F.1 · p. 23, Figure 11 · p. 23

F.1 TRANSFORMATION OPERATORS AND ACTION VALIDITY

  • ECT’s transformed actions are kinematically valid for the transformed scenes, not symmetry images of the original actions or an action-equivariance constraint.For Spatial y-mirror pairs, translation differs from the reflected command by a median of about 0.1 on the [−1, 1] action scale.Source: §F.1 · p. 23
  • ECT construction aims for executable labels and visually distinguishable scenes, and Object diagnostics exclude transformations that reduce nominal competence.X-mirror and all-mirror variants reduce Object competence; Y-mirror and shifted variants preserve ID, Semantic, and Object success, but per-suite Object Swap remains low.Source: §F.1 · p. 24, Table 19 · p. 24, Table 20 · p. 25
  • ECT pairs use the same instruction with different scene-dependent actions, and the loss trains each original with its counterpart in the same update.ECT data enter the ordinary imitation-learning pipeline; the method does not construct counterfactuals for Task perturbations, so Task gains are evidence of transfer beyond its constructed scene changes.Source: §F.1 · p. 25

F.2 ECT COUNTERFACTUALS ARE BUILT INDEPENDENTLY OF LIBERO-PRO EVALUATION

ECT training layouts are constructed independently of LIBERO-PRO evaluation layouts, and layout-distance checks show the Swap states remain outside the ID-scale neighborhood of transformed demonstrations. Across π0.5 and GR00T-N1.7, ECT improves Swap performance; data-scale results indicate that more demonstrations of the original kind do not close the counterfactual gap.

F.2 ECT COUNTERFACTUALS ARE BUILT INDEPENDENTLY OF LIBERO-PRO EVALUATION

  • ECT adds valid scene-and-action transformations to training demonstrations, while LIBERO-PRO evaluation perturbations are generated independently by rewriting BDDL fields.The whole-layout comparison finds no Swap state within the ID reference scale of transformed layouts; the smallest distance is 6.4 cm, and median distance exceeds that of original layouts in every suite.Source: §F.2 · p. 25, Table 21 · p. 25

G ECT RESULTS ON LIBERO

  • Under the Frozen-LM recipe, ECT data plus ECT loss gives the highest Swap success on three suites, while ECT data alone does so on Object; ID and Semantic success stay close to their references.On Spatial, Task success is lower under Frozen-LM than full fine-tuning for both Standard and ECT, a difference associated with the recipe rather than a unique ECT effect.Source: G ECT RESULTS ON LIBERO · p. 26

G.2 GR00T-N1.7 ECT RESULTS

  • Four-suite GR00T-N1.7 ECT raises Swap on every suite while staying within 3 points on ID and lowering Task on no suite, though four-suite training may contribute to its gain over official per-suite checkpoints.The four-suite GR00T model remains below Frozen-LM π0.5 ECT on Swap in all suites, and the recipes train different modules.Source: §G.2 · p. 26

G.3 DECOMPOSITION ACROSS SEEDS AND SUITES

  • Across suites, ECT data account for most Swap improvement; pairing adds 2 to 5 points on Spatial, Goal, and Long, has no consistent Task effect, and adds no Object gain beyond ECT data alone.Task gains over Standard come from ECT data, since the paired examples share an instruction and differ only in scene.Source: §G.3 · p. 27

G.4 DATA-SCALE SWEEP

  • With 25%, 50%, or 100% of the original demonstrations, every Swap and Task cell remains at least 34.6 points below ID; Swap barely moves on three suites and falls on Object as data increase.The sweep uses the same base model, Frozen-LM recipe, and 30k steps, with one training seed per point.Source: §G.4 · p. 27

G.5 ECT UNDER RL POST-TRAINING

After PPO post-training, both ECT models retain Swap advantages over Standard across suites, although RL lowers Swap on Long. The appendix extends evaluation to physical UR5e and CALVIN settings, reports uncertainty and behavioral analyses, and identifies remaining scope and optimization limitations.

G.5 ECT UNDER RL POST-TRAINING

  • After 40 PPO epochs, both ECT models still lead Standard on Swap in every suite, but RL lowers Swap on Long and narrows their gains.These results use one seed per arm; nuisance cells remain on par with Standard.Source: §G.5 · p. 28, Table 26 · p. 28

H.1 REAL-ROBOT UR5E STUDY

  • The UR5e study uses 400 demonstrations per model, with ECT spreading them across four physically demonstrated layouts and pairing same-object, same-grid-position examples from different layouts.The counterfactual evaluations include unseen positions and Swap configurations withheld from each ECT companion model.Source: §H.1 · p. 29, Figure 13 · p. 29

H.2 CALVIN ABC→D

  • CALVIN scenes A, B, and C provide same-instruction demonstrations with different scene-dependent actions; the reported natural-pair gain does not require newly constructed examples.Constructed examples are replay-checked for task success, and the construction comparison uses different coverage budgets.Source: §H.2 · pp. 29–30

I STATISTICAL UNCERTAINTY AND STABILITY

  • Uncertainty estimates cover episode-level variation for fixed checkpoints, while the three-seed ECT stability summary uses seed means and sample standard deviations; hierarchical-bootstrap intervals are descriptive.Rollout intervals do not capture training variability, and only three independent seeds are available for the full π0.5 method and Frozen-LM comparisons.Source: I STATISTICAL UNCERTAINTY AND STABILITY · p. 30, Table 28 · p. 30, K LIMITATIONS · p. 31

J RELATION TO OTHER ACCOUNTS OF VLA FAILURE

  • On Goal, altered instructions select other demonstrated behaviors, and scene-matched prefix readouts recover the executed task in 34/45 cases versus 22/45 for surface wording.A displaced-object stress test shows visual tracking after the policy selects the wrong target, distinguishing familiar-behavior selection from unstructured collapse.Source: J RELATION TO OTHER ACCOUNTS OF VLA… · p. 31

K LIMITATIONS

  • Swap and Task remain below ID after ECT in every suite, and pairing is not uniformly beneficial; its optimization mechanism remains open because multiple sampler changes are bundled together.The paired sampler jointly changes counterpart co-presentation, the balance of original and counterpart examples, and shared flow-matching noise and time.Source: K LIMITATIONS · p. 31
  • ECT requires action-valid, scene-dependent alternatives, and evaluation covers task-specific fine-tuning rather than generalist use, contact-rich manipulation, mobile manipulation, or navigation.Several comparisons use one training seed, and the data-scale sweep spans only a fourfold range at fixed steps.Source: K LIMITATIONS · p. 31

Related papers

Loading 2609.39971v1…