acceptodds
Under review as a conference paper at ICLR 2027

RARA: Robust Policy Fine-Tuning through Retrieval-Anchored Representation Alignment for Robotic Manipulation

Abstract

Pre-trained robot policies, such as} Vision-Language-Action (VLA) models, provide a strong foundation for generalist robotic manipulation, but adapting them to new tasks with only a few demonstrations remains challenging. Directly fine-tuning a pre-trained policy often achieves strong in-distribution performance, yet it can distort the representations learned during pre-training and weaken the model's generalization ability under out-of-distribution shifts. We propose Retrieval-Anchored Representation Alignment (RARA), a two-stage adaptation method, to address this issue. First, RARA retrieves representation features from pretraining with similar task intent and motion, and aligns the policy's representations of the target demonstrations with the pre-trained representations of the retrieved demonstrations. Second, RARA fine-tunes the policy on the target demonstrations while limiting how much its representations of the retrieved demonstrations change during fine-tuning. By reducing the representation mismatch before action learning and preserving pre-trained knowledge during adaptation, RARA limits unnecessary representation drift during fine-tuning. Experiments with Diffusion Policy and GR00T on 13 tasks in RoboTwin, RoboCasa, and on a real robot show that RARA improves the average success of fine-tuning by 9.5 points in-distribution and 13.6 points out-of-distribution, and outperforms all compared baselines.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Imitation learning, Learning representations for robotic perception and control

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_maBSOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Retrieval-Anchored Representation Alignment (RARA), a two-stage adaptation method for fine-tuning pre-trained robot policies while preserving their generalization ability. RARA first aligns the representations of the target demonstrations with those of retrieved demonstrations from pretraining, and then fine-tunes the policy on the target demonstrations while limiting representation drift. The method is evaluated on 13 tasks in RoboTwin, RoboCasa, and a real robot, showing clear improvements in success rates both in-distribution and out-of-distribution compared to baseline methods.

Critical strengths
  1. The method offers a well-motivated way to reuse task-relevant pre-training experience without directly imposing source-task action supervision. The paper is also well-presented and clearly written, making it easy to follow the proposed method.

  2. The empirical evaluation covers two policy architectures, two simulation environments, and real-world manipulation across 13 tasks. RARA improves average ID and OOD success in all three settings, with larger gains under OOD conditions.

Critical weaknesses
  1. The evaluation does not clearly establish robustness to shifts beyond pre-training coverage. In RoboTwin, pre-training already includes broad variation in backgrounds, clutter, and object positions. The results therefore support recovering pre-trained generalization, but leave robustness to previously unseen shifts unresolved.

  2. The contribution of the retrieval design is not fully isolated. STRAP differs in both retrieval and training objectives, making it difficult to attribute the performance gap. Comparing alternative retrieval mechanisms, such as Behavior Retrieval [1] or Importance Weighted Retrieval [2], under the same RARA objectives would clarify this contribution. The restriction of negatives to instruction-filtered demonstrations also lacks comparison with cross-task negative sampling.

[1] Du, M., Nair, S., Sadigh, D., Finn, C. "Behavior Retrieval: Few-Shot Imitation Learning by Querying Unlabeled Datasets." RSS 2023.

[2] Xie, A., et al. "Data Retrieval with Importance Weights for Few-Shot Imitation Learning." CoRL 2025.

Questions for authors
  1. Why does RETAIN perform substantially worse than FT, particularly on RoboCasa? How was the interpolation coefficient selected, and which parameters were merged?

  2. What is the effect of removing only the Stage-1 anchor loss, and how sensitive is RARA to (\lambda_1) and (\lambda_2)?

4 comments

We thank the reviewer for the comments. We uploaded a revised paper in which all changes are marked in blue.

W1: Robustness to shifts beyond pre-training coverage

We agree. The RoboTwin pre-training data cover a broad range of backgrounds, clutter and object positions, and the OOD backgrounds can overlap with them. What is new in RoboTwin is the target task and its objects, which the policy sees only in 50 demonstrations with a narrow layout. RoboTwin therefore tests whether fine-tuning on a new task keeps the generalization of the pre-trained policy, and the revision now states this (Appx. M).

Shifts that the pre-training data do not cover are tested in RoboCasa, where the policy is pre-trained on real-world DROID data and adapted to simulated kitchens. RARA improves OOD success over FT by 11.2 points there (Table 2), and by 19.2 points on the real robot (Fig. 5).

W2: Isolating the retrieval design

We agree that STRAP differs from RARA in both retrieval and training objective. The table below keeps the two RARA objectives fixed and changes only the retrieval or the negatives (Appx. O of the revision). With Behavior Retrieval, RARA stays close to the original (2.5 points lower in ID and 4.1 in OOD) and well above FT (67.4 / 47.3). The gain therefore comes mainly from the two objectives, and a simple retrieval by instruction and action similarity, which needs no additional model, is sufficient. A likely reason for the small drop is that the embedding of Behavior Retrieval is trained only on the pre-training data, so its image part is less reliable for the new target objects, whereas the instruction and the actions do not depend on their appearance.

Replacing the negatives of RARA with frames from other tasks gives almost the same result. Unlike standard contrastive learning, RARA does not need negatives to prevent collapse, because the target frames are aligned to a pre-trained representation that the anchor keeps fixed, and frames of other tasks are already far away in this representation. The positives matter more: with random Stage-1 pairs, RARA falls below FT (Table 3).

Success rate (%) on RoboTwin, averaged over the four tasks. All rows use the two RARA objectives:

MethodIDOOD
RARA76.058.3
RARA with Behavior Retrieval (new)73.554.2
RARA with cross-task negatives (new)74.257.3
RARA with random pre-training data (Table 3)50.334.9
RARA with random Stage-1 pairs (Table 3)65.340.6

Q1: Why RETAIN is worse than FT

In our experiments, how well weight merging works depends on how far the target task is from the pre-training data (Appx. H of the revision). In the original setting of RETAIN, from LIBERO pre-training to new LIBERO tasks, the two are similar, and merging also works well in our test. On RoboTwin, where the pre-training and target data share the simulator, robot and cameras, RETAIN keeps about 70% of the success of FT. On RoboCasa, where the policy is pre-trained on real-world DROID data and adapted to simulated kitchens, it keeps less than half. A likely explanation is that this larger gap requires a much larger change of the weights, so the interpolated weights lie between two solutions and preserve neither behavior. RARA constrains the representation of the retrieved demonstrations instead of the weights, so the policy can still change as much as the target task needs.

For the coefficient, we follow the released implementation and interpolate all parameters with one coefficient α\alpha. We tested α∈{0.25,0.5,0.75}\alpha \in \lbrace 0.25, 0.5, 0.75 \rbrace, the values that the original paper uses on DROID, and report α=0.75\alpha = 0.75, which is the best of the three. Merging only the visual encoder or only the remaining parameters does not improve over merging all parameters (Appx. K of the revision).

Q2: Stage-1 anchor loss and sensitivity to the loss weights

Without the Stage-1 anchor, the contrastive loss can also be reduced by moving the retrieved representation, so the target frames are no longer aligned with a fixed pre-trained representation. The result agrees with this (table below, and Appx. P of the revision): RARA drops from 76.0 / 58.3 to 72.4 / 44.1, so the OOD gain is lost, and the result is similar to Stage 2 without Stage 1.

RARA is not sensitive to the loss weights. When λ1=λ2\lambda_1 = \lambda_2 changes from 0.1 to 10, RARA stays 5.6 to 8.6 points above FT in ID and 6.0 to 11.0 points in OOD. We train both stages until the losses converge, and the weights then have little effect on the result.

Success rate (%) on RoboTwin, averaged over the four tasks:

MethodIDOOD
RARA (λ1=λ2=1\lambda_1 = \lambda_2 = 1)76.0 ± 1.058.3 ± 1.2
Without the Stage-1 anchor (λ1=0\lambda_1 = 0) (new)72.4 ± 1.544.1 ± 2.2
λ1=λ2=0.1\lambda_1 = \lambda_2 = 0.1 (new)75.0 ± 1.857.9 ± 1.6
λ1=λ2=10\lambda_1 = \lambda_2 = 10 (new)73.0 ± 0.953.3 ± 1.8
Stage 2 only, no Stage 1 (Table 3)73.0 ± 3.246.2 ± 4.1
FT67.4 ± 1.447.3 ± 2.4

We hope that these responses address the reviewer's concerns, and we are happy to answer any further questions.

Thank you for the additional experimental results provided in the authors' response. I will maintain my original score.

Thank you for reading our response and the additional experiments, and for your positive assessment of our work. Your comments helped us improve the paper, in particular the comparison of retrieval designs and the analysis of the Stage 1 anchor and loss weights.

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

The paper looks at a familiar problem: you fine-tune a pre-trained policy on 50–100 demos, it works in the demo scene, and then it breaks when the camera or background changes. RARA tries to fix this by reusing relevant pre-training data. It retrieves pre-training demos with similar instructions (CLIP text similarity), and matches frames by DTW over action chunks. Stage 1 then pulls the target frames' visual features toward the pre-trained features of the matched retrieved frames, using a contrastive loss. Stage 2 is normal BC on the target data, plus a loss that keeps the features of the retrieved frames close to the pre-trained model. The retrieved actions are never used for BC.

Critical strengths

The problem is real and the method is simple. It needs no architecture change, and every anchored layer in Fig. 6 beats FT. The averaged results are consistent across two policies and three settings, and the OOD gain is larger than the ID gain in each. I checked the per-task numbers in Table 2, and the headline averages (+9.5 ID / +13.6 OOD) are computed correctly. The observation that retrieved data works better as a feature anchor than as extra BC data (Co-FT in Stage 2 hurts all four ablation columns) is the most interesting part of the paper.

Critical weaknesses

The obvious alternative explanation is not ruled out, even though it seems to have been run. RARA mainly keeps the visual features close to the pre-trained encoder. Appx. D.2 mentions a "frozen reference" (FT with the anchored layers frozen) used in RQ1. It does not appear in Table 1, and its success rate is not reported anywhere. In the RoboTwin setup I would expect this baseline to be strong, because the pre-trained encoder has already seen the OOD backgrounds (next point). The ablations don't settle it either. Stage 2 alone, which is essentially feature distillation on the retrieved data, does not improve OOD over FT (46.2 vs. 47.3), and neither does Stage 1 alone (43.2). The full method's +11 points therefore comes from an interaction between two components that individually don't help, and with three seeds I am not convinced this is robust.

The experimental design favors the method, and several parts are under-specified.

  • In RoboTwin, the pre-training data already cover 100 table textures, clutter and the full table layout (Appx. D.2). The "OOD" test widens exactly these factors. So OOD here means OOD for the target data but in-distribution for the pre-training data, which is the case where anchoring to pre-trained features should help most. Appx. F confirms that RARA can be up to 20 points worse than FT when this assumption fails.
  • Retrieval in RoboTwin is nearly trivial: about five skills with templated instructions, and the pick-and-place group contains a single skill. That makes the "random pre-training data" ablation hard to interpret. If the pick-and-place targets were pre-trained only on that group, a random subset should be almost the same as the retrieved one, yet it is 25 points worse.
  • For GR00T, which covers 9 of the 13 tasks, no RARA or baseline hyperparameters are given.
  • The real-robot std seems to come from re-evaluating a single trained policy (Appx. D.1), not from training seeds, which contradicts §5.1.
  • Table 5's gains (e.g., +10.3 on Click Alarm Clock) don't match Table 2 (+0.0 ID / +11.8 OOD).

Taken together, I can't tell how much of the improvement would survive a stronger baseline and a less favorable setup.

Questions for authors
  1. Please report the success rate of the frozen-encoder FT model mentioned in Appx. D.2, on all settings. Also report feature distillation toward the policy's own pre-trained encoder on the target frames, without retrieval or Stage 1. These two results matter most for my score.
  2. In RoboTwin, was each target's policy pre-trained on its own group or on all ~2k demos together? How was the "random pre-training data" subset drawn? Do the OOD backgrounds and textures overlap with those in the pre-training set?
  3. Can you show results where the OOD factors are not covered by the pre-training data? For example, held-out textures in RoboTwin, or reporting separately the RoboCasa OOD factors that differ most from DROID.
  4. On the real robot, are the three runs three training seeds or three evaluations of the same policy? Why are Co-FT and RETAIN missing there?
  5. How do the numbers in Table 5 relate to Table 2?
4 comments

We thank the reviewer for the comments. We uploaded a revised paper in which all changes are marked in blue.

W1 / Q1: Frozen-encoder baseline and feature distillation on target frames

We ran both baselines, and neither explains the gain (table below, and Appx. L of the revision). Freezing the visual encoder lowers success below FT in both simulated settings, by 5.6 (ID) and 8.2 (OOD) points on RoboTwin and by 25.7 and 16.3 points on RoboCasa. Even on RoboTwin, where the pre-trained encoder has seen the OOD backgrounds, a frozen encoder cannot adapt its features to the new target objects. VLM4VLA (Zhang et al., 2026) also reports a large drop in success when the vision encoder of a VLA is frozen. RARA with a frozen encoder is also far below RARA, so RARA does not work by keeping the encoder fixed.

Feature distillation toward the policy's own pre-trained encoder on the target frames, without retrieval or Stage 1, gives no gain either: it reaches 31.3% ID success on Move Pill Bottle, compared with 32.9% for FT and 51.6% for RARA. A penalty on the target frames constrains exactly the frames whose representation must change to learn the new task, whereas RARA anchors the retrieved pre-training frames and leaves the target representation free to move.

Task-average success rate (%), given as ID / OOD, on RoboTwin (Diffusion Policy) and RoboCasa (GR00T N1.7):

MethodRoboTwinRoboCasa
FT67.4 / 47.369.2 / 44.3
FT, frozen visual encoder (new)61.8 / 39.143.5 / 28.0
RARA76.0 / 58.374.0 / 55.5
RARA, frozen visual encoder (new)64.6 / 44.042.9 / 33.1

W2: Each stage alone does not improve OOD success

The two stages are designed to work together (§5.5 of the revision). Stage 1 aligns the target representation with the pre-trained representation of the retrieved demonstrations, and the anchor in Stage 2 keeps this representation in place during fine-tuning. With Stage 1 only, fine-tuning moves the representation again, and the alignment is lost. With Stage 2 only, the pre-trained representation is preserved, but the target frames are not aligned with it, so the policy cannot reuse it in OOD conditions.

The gain of the full method is consistent. On RoboTwin, RARA is above FT in OOD success on all four tasks, and the average gain of 11.0 points is several times the standard deviation over seeds (1.2 for RARA and 2.4 for FT). A similar gain appears with a different policy and different tasks: 11.2 OOD points on five RoboCasa tasks and 19.2 points on four real-robot tasks. On Open Blender Lid in RoboCasa, each stage alone already gives part of the OOD gain (52.2 and 48.0 vs. 44.7 for FT, Table 3).

W3a / Q2 / Q3: OOD factors covered by the pre-training data

To Q2: in RoboTwin, one policy is pre-trained on all 2k demonstrations, and all four target tasks start from it. The "random pre-training data" ablation draws its demonstrations uniformly at random from these 2k demonstrations. The OOD backgrounds are sampled from the whole RoboTwin texture library, so they can overlap with the 100 textures of the pre-training data (Appx. M of the revision).

We therefore agree with the reviewer's reading: in RoboTwin, the target task and its objects are new, but the backgrounds, clutter and layouts are largely covered by the pre-training data. This is the case that robust fine-tuning addresses, which is to keep the generalization of the pre-trained policy while learning a new task. To Q3: OOD factors that the pre-training data do not cover are tested in RoboCasa, where the policy is pre-trained on real-world DROID data and adapted to simulated kitchens. RARA improves OOD success over FT by 11.2 points there (Table 2), and by 19.2 points on the real robot (Fig. 5). The loss of up to 20 points in Appx. G (Appx. F in the submission) has a different cause: the pre-training data there do not contain the skill of the target task.

W3b: Retrieval in RoboTwin and the random pre-training data ablation

This expectation would hold if each group were pre-trained separately. However, one policy is pre-trained on all 2k demonstrations, which cover five skills, and the random subset is drawn from all of them (see Q2 above). For a pick-and-place target, it therefore contains many demonstrations of other skills, and anchoring a different skill hurts (Appx. G of the revision).

We agree that retrieval is easy in RoboTwin. The ablation there shows that the choice of the anchored data matters, not that retrieval is hard. Retrieval is harder with DROID, which is large and has free-form instructions, and random pre-training data also lower RARA on Open Blender Lid from 80.7 / 56.7 to 67.7 / 44.2 (Table 3).

W3c: GR00T hyperparameters

We agree that these settings were missing, and Appx. N of the revision now lists them (Table 13). In short, FT, all baselines and Stage 2 of RARA use the official fine-tuning script with its default optimizer settings (learning rate 10−410^{-4}, weight decay 10−510^{-5}, warm-up ratio 0.05) and batch size 128 for 300 epochs. RARA anchors the output of the vision encoder with λ1=λ2=0.1\lambda_1 = \lambda_2 = 0.1, and Stage 1 runs for 3,000 steps with learning rate 3×10−53 \times 10^{-5}. Co-FT and STRAP mix target and pre-training samples 1:1.

W3d / Q4: Real-robot runs and missing baselines

On the real robot, the three runs are three training seeds: each method is trained three times, and each trained policy is evaluated for 10 trials per task and setting. The sentence in Appx. D.1 was misleading, and the revision corrects it together with §5.1.

Co-FT and RETAIN are the weakest baselines in simulation. One method takes three training runs and 240 real-robot trials, so we ran FT and the two strongest baselines. The real-robot setting uses the same policy and pre-training data as RoboCasa, where STRAP and BlindVLA have the highest OOD success among the baselines (48.1 and 46.8), Co-FT is below FT (67.1 / 43.7 vs. 69.2 / 44.3), and RETAIN is far below it (31.1 / 14.8). Evaluating a subset of the simulation baselines on the real robot is common practice: STRAP (Memmel et al., "STRAP: Robot Sub-Trajectory Retrieval for Augmented Policy Learning", ICLR 2025) compares with seven baselines in simulation and with three of them on the real robot, and RETAIN (Yadav et al., "Robust Finetuning of Vision-Language-Action Robot Policies via Parameter Merging", ICLR 2026) omits two of its baselines, Scratch and Freeze-FT, on the real robot.

W3e / Q5: Table 5 vs. Table 2

The two tables use different pre-trained policies and report different quantities. In Table 2, one policy is pre-trained on all 2k demonstrations. In Table 5, each row uses a separate policy pre-trained only on the named data, and each number is the gain of RARA over FT averaged over ID and OOD. For Click Alarm Clock, Table 2 gives +0.0 (ID) and +11.8 (OOD), which is +5.9 on average, and Table 5 gives +10.3 with a policy pre-trained only on Click Bell. The caption of Table 5 in the revision now states this.

We hope that these responses address the reviewer's concerns, and we are happy to answer any further questions.

Thank you for the additional experiments. The frozen-encoder results (Table 11) and especially the skill-removal experiment (Table 12) address my main alternative explanation: the gain is not just from keeping visual features near their pre-trained values. The clarifications on the RoboTwin pre-training setup, GR00T settings, real-robot seeds, and Table 5 also resolve my other questions. I am raising my score to 3 and soundness to 2.

Thank you for reading our response and the additional experiments, and for raising your score. We appreciate your careful comments, which helped us improve the paper, in particular the frozen encoder baseline, the skill removal experiment, and the clarifications on the RoboTwin setup and the real robot runs.

Reviewer_utdsOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces RARA (Retrieval-Anchored Representation Alignment), a two-stage method designed to tackle a key challenge in robotic learning: the loss of out-of-distribution (OOD) generalization when fine-tuning pre-trained policies on a small set of target demonstrations. To prevent representation drift and preserve foundational knowledge, RARA operates in two stages:

Stage 1: The framework first retrieves relevant pre-trained demonstrations based on language similarity. It then constructs positive and negative pairs using the Dynamic Time Warping (DTW) distance between action segments, aligning the representations of target demonstrations with those from pre-training.

Stage 2 : The policy undergoes behavioral cloning on the target task. Simultaneously, a regularization constraint enforces that representations of the retrieved demonstrations stay close to those of a frozen pre-trained model, safeguarding pre-trained knowledge from being overwritten.

The authors evaluate RARA across 13 tasks involving two policy architectures (Diffusion Policy and GR00T), two simulation benchmarks, and real-world robot hardware. Compared to standard fine-tuning, RARA achieves average success rate improvements of 9.5 percentage points in-distribution (ID) and 13.6 percentage points out-of-distribution (OOD).

Critical strengths
  1. Preserving pre-trained benefits while adapting to new tasks is a crucial problem for robotic foundation models. RARA introduces no additional inference-time modules, which is a practical advantage for real-world deployment.

  2. Evaluation across two policy types, two simulation benchmark, and real hardware is more convincing than validation on a single benchmark alone. In Table 2, RARA achieves the highest average ID/OOD task success rates.

Critical weaknesses
  1. Lack of comparison with LoRA. LoRA serves as an important parameter-efficient baseline in few-shot fine-tuning. The current results make it unclear whether RARA's gains significantly outperform simple LoRA, or whether RARA can provide further improvements on top of LoRA.

  2. RARA additionally incorporates 3,000 steps of representation alignment, yet the paper claims identical training budgets across methods. It is necessary to clarify the total update steps, target sample exposure count, and actual computational costs.

  3. While CKA, effective rank, and feature shift provide auxiliary evidence, they do not directly prove that performance gains stem from beneficial skill transfer or knowledge retention.

Questions for authors
  1. How do RARA's ID/OOD success rates and training costs compare with LoRA? Can combining LoRA + RARA still yield consistent gains?

  2. Are the 3,000 steps in Stage 1 counted toward the total training budget? Under the same computational budget or target sample exposure count, how does standard fine-tuning (FT) perform with extended training?

  3. How are the teacher, alignment layers, projector, and loss configured for BlindVLA? Has RETAIN been tested with blending coefficients closer to the fine-tuned model and module-wise merging?

3 comments

We thank the reviewer for the comments. We uploaded a revised paper in which all changes are marked in blue.

W1 / Q1: Comparison with LoRA

We added LoRA and LoRA + RARA on the four RoboTwin tasks (table below; settings and per-task results are in Appx. I of the revision). LoRA uses rank 16 and trains 2.8% of the parameters, and the two new rows use 50 rollouts per seed and setting. LoRA alone is below full fine-tuning (FT) in ID success and close to it in OOD success, and RARA is above LoRA by 13.7 (ID) and 12.0 (OOD) points. Combining the two still helps: adding RARA to LoRA raises the task average by 7.0 points in ID and 3.4 points in OOD. In training cost, RARA adds the same short Stage 1 and anchor loss to LoRA as to full fine-tuning (see W2 / Q2).

Success rate (%) on RoboTwin, averaged over the four tasks:

MethodIDOOD
FT67.4 ± 1.447.3 ± 2.4
RARA76.0 ± 1.058.3 ± 1.2
LoRA (new)62.3 ± 1.346.3 ± 3.1
LoRA + RARA (new)69.3 ± 1.049.7 ± 1.5

W2 / Q2: Training budget of Stage 1

Stage 1 is not counted in the fine-tuning budget. It is an additional step that only RARA runs, and §5.1 of the revision now states this. FT and Stage 2 of RARA use the same schedule (15k to 50k steps depending on the task), and the 3,000 steps of Stage 1 cost about 5–10% of it. In target sample exposure, RARA sees fewer target samples than FT, because half of each Stage-2 batch holds retrieved frames (Appx. J of the revision).

This additional computation does not explain the gain. On RoboCasa, we trained FT and RARA for 5k more steps, which is more than the whole of Stage 1, and neither method improves. FT with the longer schedule remains 5.9 (ID) and 12.2 (OOD) points below RARA with the original one.

Success rate (%) on RoboCasa, averaged over the five tasks:

MethodIDOOD
FT69.2 ± 2.544.3 ± 2.0
FT, 5k more steps (new)68.1 ± 3.143.3 ± 2.6
RARA74.0 ± 1.455.5 ± 4.8
RARA, 5k more steps (new)73.8 ± 1.855.2 ± 5.1

W3: Representation metrics as evidence

We agree that representation metrics alone are not causal evidence, and the revision now describes them as descriptive (§5.2). The causal evidence comes from three experiments that change one factor at a time (Appx. G and Appx. L of the revision). First, we change only the pre-training data, from pick-and-place demonstrations to Click Bell demonstrations. Anchoring helps when the pre-training data contain the skill of the target task and hurts otherwise (Move Pill Bottle: +18.8 vs. −6.7-6.7 points; Place Container: +4.9 vs. −20.0-20.0). Second, the same penalty applied to the target frames instead of the retrieved frames gives no gain on Move Pill Bottle (31.3 vs. 32.9 for FT, ID), while anchoring the retrieved frames gives +19 points (50.3 vs. 31.5 over five seeds). Third, when the pick-and-place demonstrations are removed from the pre-training data and RARA anchors the same retrieved demonstrations, the gain turns into a loss (ID: from +18.7 to −4.6-4.6 on Move Pill Bottle, and from +4.7 to −15.3-15.3 on Place Container). The gain therefore depends on what the pre-trained policy has learned from these demonstrations, and not only on the images.

Q3: Configuration of BlindVLA and RETAIN

BlindVLA was proposed for OpenVLA. For a fair comparison, we apply it to the same base policy as all other methods. As in the original paper, it uses a frozen projector and a patch-wise cosine loss. The teacher is a frozen DINOv2-B/14, aligned with ResNet layer 3, for the Diffusion Policy, and a frozen C-RADIOv3-L, aligned with the output of the vision encoder, for GR00T N1.7 (Appx. N of the revision). We also tried the original OpenVLA setting in RoboCasa, but fine-tuned OpenVLA is far below fine-tuned GR00T N1.7, so that comparison would mainly reflect the base model.

For RETAIN, we follow the released implementation, which interpolates all parameters. We tested α∈{0.25,0.5,0.75}\alpha \in \lbrace 0.25, 0.5, 0.75 \rbrace, the values that the original paper uses on DROID, and Table 2 reports α=0.75\alpha = 0.75, which is the closest to the fine-tuned policy and the best of the three (table below, and Appx. K of the revision). Success increases with α\alpha, and FT (α=1\alpha = 1) is the best, so a coefficient closer to the fine-tuned policy does not bring RETAIN above FT. Module-wise merging does not help either.

Success rate (%) of RETAIN on RoboTwin, averaged over the four tasks:

CoefficientIDOOD
α=0.25\alpha = 0.259.8 ± 8.03.4 ± 3.1
α=0.5\alpha = 0.538.9 ± 7.924.5 ± 6.2
α=0.75\alpha = 0.75 (Table 2)47.0 ± 6.734.2 ± 2.3
α=0.75\alpha = 0.75, visual encoder only45.4 ± 7.233.1 ± 1.9
α=0.75\alpha = 0.75, all but the visual encoder39.8 ± 8.330.1 ± 2.7
α=1\alpha = 1 (FT)67.4 ± 1.447.3 ± 2.4

We hope that these responses address the reviewer's concerns, and we are happy to answer any further questions.

Thank you for the response and additional experiments. The new results address the missing LoRA baseline and partially alleviate my concerns about the training budget. Using relevant pre-training demonstrations both to align representations across domains and to limit representation drift during fine-tuning is useful, although the overall method remains incremental. Overall, I consider the paper borderline but lean toward weak acceptance. I have therefore raised my score to 3.

Thank you for reading our response and the additional experiments, and for raising your score. We appreciate your comments, which helped us improve the paper, in particular the LoRA comparison and the description of the training budget.

Comments on the paper

1 comment

Dear Reviewers,

As the discussion period is coming to a close, please take a moment to review the authors’ responses and revised manuscript. Kindly assess whether your concerns have been adequately addressed or if any questions remain, and update your scores accordingly, if necessary.

Thank you for your time and thoughtful evaluations.

Best regards, AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.