acceptodds
Under review as a conference paper at ICLR 2027

Change2Act: Object-Centric Visual Reasoning for Image-Goal Manipulation

Abstract

Robot manipulation is a fundamental capability for embodied intelligence in real-world environments. Recent generalist robot policies typically specify task goals through natural language and generate continuous actions from visual observations. However, specifying tasks in complex scenes requires precise instructions about objects and their spatial relationships, and writing such instructions demands substantial human effort. To this end, we introduce Image-Goal Manipulation (IGM), where a goal image directly specifies the desired environment state and the policy generates continuous manipulation actions from the current visual observation to realize that state. IGM poses two key challenges. First, the policy must identify task-relevant objects and their desired state transitions despite substantial visual differences between the current observation and goal image. Second, these transitions must be converted into effective action conditions that guide object interaction and continuous control. To address these challenges, we propose Change2Act, a framework that connects explicit visual change reasoning with lightweight policy adaptation. Specifically, we design an Object-Centric Change Reasoning module that establishes semantic correspondences between objects in the initial and goal images and encodes their spatial differences into a structured change descriptor. We further design a Change-Guided Action Conditioning module that maps this descriptor into compact goal tokens and combines them with the original goal image to guide a pretrained robot policy. A goal-role embedding distinguishes the desired state from current observations. Experiments demonstrate that Change2Act achieves more accurate offline action predictions than the compared baselines while requiring fewer trainable parameters.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Image-Goal Manipulation, Object-Centric Change Reasoning

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_9prgOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes Change2Act, a method that uses image-goals as prompts for VLAs through an object-centric reasoning module. It operates by encoding spatial semantic differences into a descriptor that is mapped to goal tokens through several pre-trained models (SAM, DINO, GroundingDINO). Together with the original goal image, a pre-trained robot policy (π0.5\pi_{0.5}) is adapted towards producing actions that reach the goal.

Critical strengths
  • The overall idea is simple and intuitive (the implementation less so).
  • Error prediction accuracy is a bit better than prior methods.
Critical weaknesses
  • The underlying assumption that text prompting requires "substantial human effort” compared to image goal is misleading. How would one acquire an image-goal in the real world? It would typically require a human to organize the objects into the desired configuration, which arguably might require more effort than just text (and the human has to complete the task before the robot). I suggest toning down the claim and provide real-world examples where goal-images are easily acquired.
  • Novelty and prior work: Goal-VLA appears to be doing something similar but on more complex tasks. There is also prior work that utilizes SAM-based masks and/or visual reasoning such as VP-VLA and Liu et al..
  • Complex masking and correspondence pipeline: to obtain a change mask, several stages are required that might require low-level tuning, such as adaptive thresholding and morphological processing. Additional filtering stages that might be brittle. The values for all of these stages are nowhere to be found.
  • Reliance on pre-trained models: the model largely relies on the quality of several pre-trained models (SAM, DINO, GroundingDINO) and heuristics tuning for filtering.
  • Limited benchmark: model is only evaluated on the saturated LIBERO dataset.
  • Evaluation: this is the most significant concern - the authors only measure action prediction error rather than task success. This is a weird choice for evaluation, and not reporting success rates raises serious concerns regarding the soundness and validity of the proposed approach (there is one case study, but definitely not enough).
  • Limitations are not discussed.
Questions for authors

See Weaknesses. I don't believe this paper is ready for publication.

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

Change2Act adapts a frozen π05 policy for image-goal-conditioned manipulation. It uses pretrained perception models and simple object matching to extract a change descriptor from initial to goal. A trainable adapter turns this into goal tokens. The experiments use LIBERO-derived demonstrations. They mostly test offline action prediction, and also include some ablations and three qualitative execution examples.

Critical strengths

The framework is straightforward to understand and requires relatively few trainable parameters.

Critical weaknesses
  1. The task formulation is not new. Manipulation conditioned on a goal image is an established setting, and the paper itself acknowledges this capability in Octo. Presenting Image-Goal Manipulation as a newly introduced formulation therefore overstates the task-level contribution.

  2. The methodological novelty is limited. The pipeline assembles pretrained SAM, Grounding DINO, and DINOv2 with image differencing, heuristic correspondence scoring, feature concatenation, and a lightweight conditioning adapter. The paper does not establish a sufficiently novel learning mechanism or technical insight beyond this integration of existing components.

  3. The main evaluation does not establish closed-loop effectiveness. Offline action MAE/RMSE and gripper agreement cannot substitute for task success rates under policy execution. The three qualitative examples against AVA-VLA are insufficient without systematic rollouts, repeated trials, and comparison against the strongest baseline, π05.

  4. The goal-image setting is overly controlled relative to the practical motivation. Goals are final frames from the same demonstrations and camera views, while the change-extraction procedure relies on aligned-view image differencing. The paper does not test independently acquired goal images, cross-demonstration references, or viewpoint changes, leaving the practical availability and robustness of the required goal inputs unresolved.

  5. The baselines do not adequately isolate the proposed contribution. The parameter-efficiency study lacks a comparable-budget π05 baseline, despite π05 being the proposed method’s backbone. A matched-backbone and matched-budget comparison with direct goal-image conditioning and a generic learned-token adapter is needed to distinguish the benefit of explicit change extraction from the adaptation strategy itself.

  6. The evaluation scale and statistical reliability are insufficiently documented.

Questions for authors

please see weakness.

Reviewer_vdVzOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary
  1. Change2Act addresses image-goal manipulation by identifying object-level changes between initial and goal images and using them to condition a pretrained robot policy.

  2. The method combines frozen detection, segmentation, and visual features with heuristic object matching, then trains a goal- token adapter and goal-role embedding on a frozen π0.5 backbone.

  3. On the LIBERO-derived IGM-Restore dataset, it improves offline action prediction with 1.649M trainable parameters; execution evidence consists of three qualitative examples.

Critical strengths
  1. Focused, parameter-efficient design. The object-change descriptor provides explicit task cues while leaving the policy backbone frozen, requiring only 1.649M trainable parameters.

  2. Consistent offline improvements. Action MAE decreases from 0.061 to 0.038 relative to the π0.5 baseline; removing the change descriptor increases it to 0.046, supporting its incremental value (Tables 1–2).

  3. Component-level analysis beyond aggregate scores. The appendix evaluates object correspondence and localization directly and includes controlled descriptor interventions, helping assess whether the proposed information is actually used.

Critical weaknesses
  1. Manipulation effectiveness is not adequately established. Evaluation centers on matching expert actions, without aggregate closed-loop success rates. Three qualitative rollouts cannot establish reliable goal achievement or robustness to execution errors.

  2. The pretrained checkpoint creates an unresolved contamination concern. The backbone is already LIBERO-pretrained, while IGM- Restore is derived from LIBERO demonstrations. Episode-level splitting of the downstream dataset does not establish separation from backbone training data; checkpoint overlap and baseline initialization need clarification.

  3. The demonstrated scope is substantially narrower than the framing. Matching assumes aligned viewpoints, uses pixel differencing, and selects a single object pair. Multi-object changes and viewpoint shifts are untested. Moreover, the paper itself acknowledges Octo’s goal-image conditioning, so “introducing IGM” overstates task novelty.

Questions for authors
  1. Can you report closed-loop success rates and final goal-state errors across the full test set, with repeated trials and uncertainty estimates, rather than only offline action errors?

  2. Which LIBERO demonstrations were used to train the π0.5 checkpoint, and do they overlap with IGM-Restore validation/test episodes? Do the π0.5 baseline and Change2Act start from exactly the same checkpoint?

  3. How does the method handle camera shifts, multiple changed objects, or no valid correspondence? Can you evaluate these cases and distinguish the contribution from existing image-goal policies?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.