acceptodds
Under review as a conference paper at ICLR 2027

StereoPolicy: Improving Robotic Manipulation Policies via Stereo Perception

Abstract

Recent advances in robot imitation learning have produced powerful visuomotor policies that manipulate diverse objects from visual inputs. However, monocular observations lack depth information, which is critical for precise manipulation in cluttered or geometrically complex scenes. Explicit depth maps and point clouds are often noisy and fragile in real-world manipulation. We introduce StereoPolicy, a visuomotor policy learning framework that directly leverages synchronized stereo image pairs to improve spatial perception without constructing explicit 3D representations. StereoPolicy processes each image with pretrained 2D vision encoders and fuses left-right features through a cross-attention-based Stereo Transformer, learning task-conditioned binocular representations. The framework integrates with diffusion-based and pretrained vision-language-action (VLA) policies, delivering consistent improvements over RGB, RGB-D, point cloud, and multi-view baselines across three simulation benchmarks and seven real-robot tabletop and bimanual mobile manipulation tasks. Our results support task-trained binocular fusion as an effective way to adapt 2D visual representations for manipulation.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Imitation learning, Robot Manipulation, Stereo Vision

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_E3EzOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents StereoPolicy, a framework that learns robotic manipulation policies directly from synchronized stereo images without explicit depth supervision or 3D reconstruction. It combines shared-weight image encoders with a Stereo Transformer that alternates self-attention and cross-attention to produce binocular features trained through action objectives. The framework is integrated into diffusion policies and pretrained VLA models, and evaluated on three simulation benchmarks and seven real-world tasks, with additional ablations, relative-depth probes, and camera-configuration studies.

Critical strengths
  1. The paper demonstrates improvements across diffusion policies and pretrained VLA models in both simulation and real-world manipulation, including comparisons with native two-image VLA inputs. This breadth supports the practical value of the approach.
  2. Duplicated-input controls and matched joint-attention comparisons support the contribution of binocular information and the fusion design. Relative-depth probes provide additional evidence that action training makes geometric information more accessible in the learned representations.
Critical weaknesses
  1. The architecture combines established components, including shared image encoders, alternating self-attention/cross-attention, positional encoding, and action supervision. While the matched joint-attention comparison supports the effectiveness of the chosen design, its methodological contribution relative to existing direct stereo-control approaches could be more clearly distinguished. Moreover, Table 3 compares a trainable StereoPolicy visual module with frozen FoundationStereo features, making it difficult to separate the effects of representation choice and task adaptation. Overall, the empirical contribution is more compelling than the architectural novelty.

  2. Table 8 shows that gripper–target relative depth becomes more linearly decodable after action training, including on held-out trajectories and spatial layouts. However, this establishes access to one task-relevant geometric quantity, leaving unclear whether the features encode broader, transferable spatial relationships and whether the policy uses the probed information to improve action selection. The evidence supports task-adapted features containing geometric information, but provides a more limited explanation of what geometry is learned and how it contributes to control.

Questions for authors
  1. In Table 3, StereoPolicy’s visual module is trained with the action loss, while the FoundationStereo feature extractor is frozen. Have the authors tried jointly fine-tuning FoundationStereo and the action policy using the same action loss? This comparison would help clarify whether the performance gap comes mainly from the different architectures or from allowing only StereoPolicy’s visual features to adapt to the task.

  2. Appendix C.4 states that the highest success rate observed during training is reported for the Robomimic and OmniGibson experiments. Are these scores reported directly from the periodic evaluations, or are the selected checkpoints evaluated again on fresh, independent episodes? Please also clarify how checkpoint selection is performed across the three training seeds. Reporting the maximum observed evaluation score directly may lead to optimistic performance estimates.

3 comments
AuthorsOpenReview ↗

Response to Reviewer E3Ez

We thank Reviewer E3Ez for their time and detailed feedback!

Q1. FoundationStereo fine-tuning and task adaptation

We thank reviewer for identifying this important distinction. To disentangle the effect of architecture from task-specific feature adaptation, we additionally experimented by jointly fine-tuning FoundationStereo and the action policy using the same action loss and show the results below (ToolHang-100). Results are reported as mean success rate and standard deviation across three training seeds.

MethodSR (%)
STEREOPOLICY-DP94 (1.9)
Frozen FS features76 (4.3)
Joint Fine-tuning FS84 (1.6)

Jointly fine-tuning decreases the gap but still underperforms StereoPolicy-DP. This additional experiment strengthens the evidence for the effectiveness of StereoPolicy’s task-trained stereo representation for manipulation-policy learning.

Q2. Checkpoint selection and independent testing

We fix the previously selected best checkpoints before testing on fresh evaluation seeds, retaining three training seeds and 100 episodes per task and seed. Please see C2 (Tables R1–R2) for the protocol and results.

W1. Contribution and prior stereo-control work

Please see C1.

W2. Interpretation of geometric probes: Table 8 shows that gripper–target relative depth becomes more linearly decodable ...

We use gripper–target relative depth as a focused diagnostic of spatial information relevant to manipulation. On held-out trajectories and layouts, action training reduces linear-probe MAE from 3.1 to 1.1 cm on ToolHang and from 4.0 to 1.5 cm on PourWater (Table 8). Matched policy ablations separately show the control benefits of stereo input and fusion (Table 7). These provide complementary evidence of learned geometric information and improved manipulation performance, without treating probe accuracy as proof of a specific causal mechanism or a general-purpose 3D representation.

We thank the reviewer again for the careful review and constructive suggestions, and welcome any further questions.

I thanks the author for the extra experiments and that solves my concerns. I therefore raise my evaluation to 3.

Thank you for taking the time to look over the additional experiments. We’re glad they addressed your concerns, and we’d be happy to answer any further questions!

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

StereoPolicy introduces a stereo perception module for robotic manipulation. It combines image features through a Stereo Transformer and learns binocular representations using the action objective, without explicit depth reconstruction or auxiliary geometric supervision. The module is integrated into diffusion policies and pretrained VLAs, including π0.5 and GR00T-N1.5. Experiments cover three simulation benchmarks and seven real-world tasks, alongside fusion ablations, geometric probes, and camera-configuration analyses.

Critical strengths
  1. Useful controlled experiments. Parameter-matched baselines, duplicated-left-image controls, joint-attention comparisons, and relative-depth probes provide evidence that the model exploits binocular information. These experiments strengthen the argument beyond simply adding a second image.

  2. Practical applicability. The module can be integrated into both diffusion policies and pretrained VLAs. Evaluation across simulation and real-world settings demonstrates practical relevance, particularly for tasks requiring precise spatial alignment.

Critical weaknesses
  1. Limited methodological novelty. The approach primarily combines established components: pretrained image encoders, multi-scale features, and self-/cross-attention fusion. Closely related stereo-policy designs already exist; for example, DextrAH-RGB uses shared image encoders and cross-view attention for visuomotor control. Learning binocular fusion through the action objective without additional geometric supervision is a useful simplification, but the manuscript does not clearly establish a substantial conceptual or architectural advance over existing formulations. This limits the significance of the methodological contribution.

  2. Limited validation of the comparative claims. The paper includes representative baselines such as DP3 and 3D Diffuser Actor, and the matched fusion ablations are helpful. However, the comparison with frozen pretrained stereo representations in Table 3 is limited to one task with a diffusion-policy backbone. It therefore remains unclear how consistently the observed advantage over this alternative holds across tasks and policy families. Additionally, only 20 real-world trials are reported per task, leaving uncertainty around smaller performance differences. It is unclear whether checkpoint selection and final testing use independent evaluations. If the same evaluations serve both purposes, reporting the highest success rate across training may produce optimistic estimates.

Questions for authors
  1. Contribution clarification: What specific architectural or learning contribution distinguishes StereoPolicy from existing stereo-attention encoders, and which experiments most directly isolate that contribution?

  2. Baseline clarification: Which FoundationStereo features and semantic components are retained in the StereoVLA-style baseline? How are feature alignment, adaptation, and training conditions configured, and what conclusions is this comparison intended to support?

  3. Evaluation clarification: Are checkpoints selected using separate validation evaluations before independent final testing? Do the main conclusions hold under this procedure?

  4. Statistical clarification: Were the real-world methods evaluated on matched initial configurations and multiple training seeds? Can uncertainty estimates be provided for the principal comparisons?

2 comments
AuthorsOpenReview ↗

Response to Reviewer XaMT

We thank Reviewer XaMT for their time and detailed feedback!

Q1&W1. Contribution and distinction from prior stereo methods

Please see C1.

Q2. FoundationStereo baseline configuration and comparison scope

We implement a StereoVLA-inspired feature-replacement baseline rather than reproducing the complete StereoVLA training procedure. Our experiments intend to showcase whether an off-the-shelf stereo-geometric prior plus pretrained semantics can replace our task-supervised visual encoder and Stereo Transformer.

Specifically, the stereoVLA-style baseline retains DINOv2, SigLIP and FoundationStereo’s frozen filtered cost volume without StereoVLA’s auxiliary depth loss. We spatially pool FS’s frozen filtered cost volume and concatenate the features channel-wise, and use a trainable projector and two-layer Transformer to produce the same concatenated 256-D input expected by the unchanged policy backbone. All other training settings are held fixed. This comparison therefore isolates whether pretrained stereo-geometric and semantic features can replace StereoPolicy’s task-supervised visual representation, rather than evaluating the full StereoVLA architecture or explicit depth supervision.

Q3. Checkpoint selection and independent evaluation

Please see C2 (Tables R1–R2) for the independent evaluation protocol and results. The overall performance advantage holds across all three simulation benchmarks.

Q4&W2. Real-world evaluation protocol and uncertainty

Please see C3.

W2. Additional comparison with frozen pre-trained stereo representations across tasks

We additionally report the following results, which extends the comparison to PourWater task with 200 demonstrations. StereoPolicy achieves 70% success versus 58% with frozen FoundationStereo features, showing that the advantage is not limited to ToolHang task.

MethodSR(%)
STEREOPOLICY-DP70
Frozen FS features58

We thank the reviewer again for the careful review and constructive suggestions, and welcome any further questions.

Dear Reviewer XaMT,

Thank you again for your time and thoughtful questions. We ran additional experiments in response and will include the results and clarifications in the revised paper. We hope our rebuttal addresses your concerns. If anything remains unclear, we’d be happy to follow up before the discussion closes!

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

The paper proposes StereoPolicy, a stereo perception module for visuomotor policies. Each image of a rectified stereo pair is encoded by a shared 2D encoder (ResNet18+FPN), with frozen DINOv2 features concatenated for external cameras. A 2-layer Stereo Transformer then alternates within-image self-attention and left-right cross-attention with 2D RoPE. The fused tokens either condition a diffusion policy trained from scratch (StereoPolicy-DP) or replace the visual embedding of a pretrained VLA (π0.5, GR00T-N1.5), which is then fine-tuned end-to-end (StereoPolicy-VLA). Baselines include RGB, RGB-D, RGBD-3DDA, point clouds (PointNet, DP3) and a two-image Multi-View baseline.

Critical strengths
  1. A broad, matched comparison of input modalities for imitation learning. With the same policy backbone and training budget, raw stereo input is compared against RGB, RGB-D, RGBD-3DDA, point-cloud and two-image baselines across three simulators and real tabletop and mobile manipulation.
  2. The controls rule out a "more pixels or parameters" explanation.
Critical weaknesses
  1. Limited novelty and an unclear relation to the closest prior work. DextrAH-RGB already encodes the left and right images with a shared, fine-tuned ResNet-18 and fuses them with a 2-layer transformer in which each view's tokens cross-attend to the other view. The distinctions claimed in Related Work (learning through the action objective, generic pretrained 2D encoders, no disparity estimation, no depth supervision) all apply to DextrAH-RGB as well. Open-TeleVision (Cheng et al., CoRL 2024, arXiv:2407.01512) feeds both stereo images to an ACT policy and reports a large drop in success when the right image is removed, which is essentially the joint-attention variant in Table 7. StereoVLA (Deng et al., RSS 2026) was published before July 17, 2026, the cutoff for contemporaneous work in the ICLR 2027 reviewer guidelines, so it should not be described as concurrent. It already adapts a VLA to stereo input by fusing FoundationStereo cost-volume features with SigLIP and DINOv2 semantic features, and it evaluates π0.5 with the left and right images as its two image inputs. This overlaps with Q2 and the two-image baseline in Table 9, so StereoPolicy-VLA should be positioned against StereoVLA as prior work. The alternating self-/cross-attention block with RoPE follows STTR, CroCo v2 and DUSt3R (Wang et al., CVPR 2024, arXiv:2312.14132), and this design source should be stated. The methodological contribution is therefore incremental, and the paper's value rests mainly on its empirical results.

  2. Several empirical claims are stronger than the evidence. (i) The "baseline ≈ 10% of distance" rule conflicts with the paper's own text and Fig. 9: the recommended 6 cm at 0.6-0.8 m gives r = 0.076-0.10, mostly outside the stated optimum [0.09, 0.13], and in Fig. 9 every baseline drops between 0.8 m and 0.9 m, so distance rather than the ratio explains most of the variation. (ii) The two "no fusion" controls disagree: Multi-View (Table 2) and "w/o StereoTransformer" (Table 5) are described almost identically yet reach 64% and 82% on ToolHang-100, so the gain attributed to the fusion design ranges from 30 to 12 points. (iii) Simulation results report the best checkpoint over training (Appendix C.4), i.e., model selection on the evaluation rollouts. On the real robot (20 trials per task), no per-task difference against RGB, RGB-D, RGBD-3DDA or Multi-View is significant, and the pooled gap to RGBD-3DDA and Multi-View (59/100 vs. 45/100) is borderline (p ≈ 0.07). The abstract's claim of consistent gains across seven real-robot tasks also overstates the scope, since the two mobile tasks are compared only against RGB. Because the contribution is mainly empirical, these issues directly weaken the paper's main claims.

Questions for authors
  1. What differs between the Multi-View baseline (Table 2, 64% on ToolHang-100) and "w/o StereoTransformer" (Table 5, 82%)? Which one is the intended no-fusion control?
  2. Do the rankings hold if all methods, including the VLAs, are evaluated at the final checkpoint rather than the best one?
  3. For Fig. 9, please give the numeric grid with seeds and an RGB baseline at each distance. Why does 10 cm at 0.89-0.99 m (r ≈ 0.10-0.11) perform poorly although it lies in the stated optimal range?
  4. How does the stereo encoder differ from DextrAH-RGB's, and how does it compare directly? Where does the 2D RoPE design attributed to Singh et al. come from, given that the public DextrAH-RGB paper does not mention positional encoding?
  5. What image resolution is used, and were the FoundationStereo-based RGB-D/3DDA policies trained on FS depth or only tested on it? Does the "StereoVLA-style" baseline include StereoVLA's semantic features and auxiliary depth loss?
3 comments
AuthorsOpenReview ↗

Response to Reviewer rbLN

We thank Reviewer rbLN for their time and detailed feedback!

Q1. Multi-View vs. w/o StereoTransformer

We thank reviewer for pointing out this confusion and will further clarify the distinction in our revision. The intended no-fusion ablation is “w/o StereoTransformer” in Table 5: it retains the visual encoder and removes the StereoTransformer. Restoring the module raises ToolHang success from 82% to 94% (+12 percentage points).

Multi-View (Table 2, 64%) is a parameter-matched baseline, where features from the two images are concatenated without DINOv2, FPN or token-level left–right attention, followed by an enlarged projection/MLP whose capacity is adjusted to match StereoPolicy. Separately, Table 7 holds the encoder configuration fixed: our fusion reaches 94% versus 82% for joint attention, and replacing the right image with a duplicate left image reduces success to 62%. These controls distinguish fusion design from the contribution of binocular input.

Q2. Final checkpoint vs. best checkpoint

Please see C2 (Tables R1–R2) for final-checkpoint and independent-test results across all three simulation benchmarks.

Q3. Stereo baseline and camera–object distance

StereoPolicy outperforms the matched-distance RGB baseline across all 25 stereo configurations by 26–43 percentage points. The table reports mean SR (%) over three training seeds; columns use the camera–object distance labels (m) from Figure 12.

Input / stereo baseline0.590.690.790.890.99
RGB5153533834
Stereo (2 cm)8890867270
Stereo (4 cm)9088887068
Stereo (6 cm)9494866464
Stereo (8 cm)9494886664
Stereo (10 cm)9188946866

StereoPolicy retains a 26–36-point advantage over RGB at 0.89–0.99 m. The RGB results indicate that distance affects both modalities; our ~10% ratio is a practical guideline for the tested 0.59–0.79 m range, rather than a distance-independent optimum.

Q4. DextrAH-RGB and the source of 2D RoPE

DextrAH-RGB uses masked cross-view attention and a learned summary token for an LSTM grasping policy, trained through teacher distillation with auxiliary object-position supervision. StereoPolicy alternates within-view self-attention and cross-view attention, learns through action objectives without auxiliary geometry supervision, and supports diffusion policies and pretrained VLAs. Table 7 directly evaluates our fusion design against joint attention under a fixed encoder configuration, showing 94% versus 82% success on ToolHang-100. This supports the chosen fusion design; it is not a direct comparison against DextrAH-RGB’s encoder. Citation of 2D RoPE: We thank the reviewer for pointing this out. Our attribution of the 2D RoPE design to Singh et al. was a citation error in writing, and we apologize for the resulting confusion. The use of 2D RoPE in StereoPolicy is an architectural choice made specifically for our Stereo Transformer to preserve the two-dimensional spatial organization of visual tokens during cross-view fusion; it is not inherited from DextrAH-RGB. We will correct the attribution to the relevant positional-encoding literature in the revision, including Heo et al. (2024).

Q5. Image resolution and FoundationStereo baseline configuration

  • Policy inputs are 224x224 resolution images.
  • The FoundationStereo-based RGB-D/3DDA policies are both trained and tested with FS depth.
  • The StereoVLA-style baseline includes StereoVLA’s semantic features. We test it as a pretrained-feature adaptation to our diffusion-policy backbone, not a reproduction of the full StereoVLA training pipeline. We intentionally omit disparity decoding and auxiliary depth loss, allowing us to isolate whether pretrained stereo-geometric representations improve policy learning under the same task-only supervision.

W1. Contributions

Please see C1 for our detailed contributions.

W2. Real-world statistics and evaluation scope

Please see C3.

The mobile experiments ask whether stereo can improve an existing RGB-pretrained VLA. The π0.5 model used here has no native depth or point-cloud input, and was not pretrained for those modalities; introducing them would require additional architecture and modality adaptation. We therefore compare against the same pretrained π0.5 with RGB observations. The seven real-world tasks serve two purposes: five tabletop tasks compare visual modalities, while two mobile tasks test integration with a pretrained VLA. The goal is a practical, low-overhead alternative to explicit 3D processing, rather than a ranking of separately pretrained VLA systems.

We thank the reviewer again for the careful review and constructive suggestions, and welcome any further questions.

The final-checkpoint and independent evaluations largely address my checkpoint-selection concern. However, the clarification that Multi-View excludes FPN and DINOv2 conflicts with the manuscript’s matched-encoder description, leaving this comparison confounded by encoder differences. Table 7 provides useful controlled evidence for fusion gains, but does not resolve this issue across the main evaluations. Given the incremental methodological contribution, limited comparisons with the closest stereo approaches, and unresolved uncertainty in the real-world results, I maintain my weak-reject recommendation.

We thank Reviewer rbLN for the follow-up, and are glad the new evaluations address the checkpoint-selection concern.

Multi-View baseline. Multi-View shares the same ResNet-18 image encoder as StereoPolicy-DP and is parameter-matched to it; we will make this wording precise in §4.1 and Appendix C.3 in the revision.

Limited comparisons with the closest stereo approaches. StereoVLA-style FoundationStereo features, whether frozen or fine-tuned, underperform StereoPolicy, and joint attention covers the Open-TeleVision-style design on multiple tasks. With six modality baselines across three simulation benchmarks and seven real-world tasks, we believe the evaluation is comprehensive; reproducing DextrAH-RGB’s RL-distillation pipeline is outside the scope of an imitation-learning study.

Limited contribution. As we discussed in C1, beyond architecture design, our contribution is a empirical study of stereo as a visual modality for manipulation with comprehensive experimental supports: three simulation benchmarks, seven real-world tasks, two policy families, six modality baselines, and controlled fusion, probing, and hardware analyses. We respectfully believe it goes well beyond what can be characterized as limited.

We respect the reviewer’s assessment and would appreciate the opportunity to address any further questions.

Comments on the paper

3 comments
AuthorsOpenReview ↗

General Repsonse

We thank the reviewers for their careful reading and constructive feedback! We highlight four strengths recognized in the reviews:

  1. Broad comparisons of visual modalities under shared policy backbones and training budgets (rbLN).
  2. Controlled evidence for the benefits of binocular input beyond simply adding images or model capacity (rbLN, XaMT, E3Ez).
  3. Practical integration with diffusion policies and pretrained VLAs, validated in simulation and real-world manipulation (XaMT, E3Ez).
  4. Relative-depth probes showing that action training makes geometric information more accessible in the learned features (E3Ez).

Below, we summarize our responses to the main concerns.

C1. Contribution and positioning to prior work. (rbLN, XaMT, E3Ez)

To our knowledge, our work is the first comprehensive evaluation of stereo as an alternative visual modality for robotic manipulation spanning both policies trained from scratch and large pretrained VLAs (π0.5 and GR00T-N1.5). StereoPolicy provides systematic evidence that stereo is a practical alternative to monocular RGB and explicit 3D inputs. Across three simulation benchmarks and seven real-world tasks, we combine several modality comparisons, controlled fusion ablations, geometric probes, and provide hardware studies that guide practical camera setup.

The controlled experiments distinguish our contribution from simply supplying another image. With the encoder configuration held fixed, StereoPolicy achieves 94% success on ToolHang-100 versus 82% for joint attention; replacing the right image with a duplicate left image reduces success to 62% (Table 7). StereoPolicy also improves over native two-image inputs for both π0.5 and GR00T-N1.5 across all three demonstration budgets (Table 9). These results support the effectiveness of task-trained stereo fusion across both policy families. Compared with DextrAH-RGB’s masked cross-view attention and learned-token readout, StereoPolicy alternates within-view self-attention and cross-view attention. It trains through the action objective without auxiliary object-position supervision and supports both diffusion policies and pretrained VLAs. Unlike StereoVLA, StereoPolicy learns stereo fusion through action supervision without pretrained stereo-matching features or auxiliary depth supervision. Open-TeleVision demonstrates the value of stereo input for ACT. In the revision, we will position StereoVLA as prior work and explicitly credit the attention designs in STTR, CroCo v2, and DUSt3R.

C2. Checkpoint selection and independent evaluation.(rbLN, XaMT)

We additionally evaluate all methods at the final checkpoint and independently re-evaluate the previously selected checkpoints. Both evaluations retain three training seeds and 100 episodes per task and training seed. Final evaluates the last checkpoint at the prescribed training endpoint. For Independent, we fix the previously selected checkpoints before testing on fresh evaluation seeds that were not used for checkpoint selection. All methods use the same fresh evaluation seeds, and no further checkpoint selection is performed using these test results. For the originally reported results, checkpoint selection was performed by first selecting the best checkpoint given one seed, then the same checkpoint is evaluated by another two different seeds.

On OmniGibson and RoboMimic, StereoPolicy leads the strongest baseline by 8.38 and 5.67 percentage points at the final checkpoint, and by 6.88 and 4.33 points in independent testing, respectively. On RoboCasa, it outperforms both input baselines for both VLA backbones at all three demonstration budgets under both evaluations. These results support the overall performance advantage beyond the best scores observed during training.

We fully appreciate reviewers’ suggestions and provide more detailed results by reporting the following additional results:

  • Table R1. Diffusion policy success rates (%).

Reported denotes the submitted-paper results. Values are equally weighted means over the eight task–demonstration settings in OmniGibson and the six in RoboMimic (submitted Table 2). Bold denotes the best mean in each benchmark and evaluation column.

BenchmarkMethodReportedFinalIndependent
OmniGibsonRGB55.3852.6352.88
OmniGibsonRGB-D55.8854.0052.75
OmniGibsonRGBD-3DDA65.3862.0062.88
OmniGibsonPCD-DP342.5041.2541.63
OmniGibsonMulti-View60.2558.8856.75
OmniGibsonStereoPolicy-DP72.5070.3869.75
RoboMimicRGB83.8382.5081.00
RoboMimicRGB-D83.8381.1779.50
RoboMimicRGBD-3DDA91.0087.6788.00
RoboMimicPCD-DP368.0063.5068.17
RoboMimicMulti-View86.3384.3382.83
RoboMimicStereoPolicy-DP94.6793.3392.33

At the task level, StereoPolicy is first or tied in all 14 DP settings at the final checkpoint and first in 13/14 in independent testing. The exception is Square-100 (84% vs. 89% for RGBD-3DDA).

  • Table R2. VLA success rates (%) on RoboCasa.

Each entry is the mean over the 24 tasks at the given demonstration budget, following submitted Table 9. Bold denotes the best result for each backbone and budget.

BackboneDemosInputReportedFinalIndependent
π0.530Monocular48.7146.7745.63
π0.530Two-image49.0146.9946.45
π0.530StereoPolicy51.7251.4551.30
π0.5100Monocular67.6467.6466.01
π0.5100Two-image68.2467.4367.30
π0.5100StereoPolicy71.5471.3970.84
π0.5300Monocular70.3170.1266.40
π0.5300Two-image70.8169.7768.92
π0.5300StereoPolicy74.4074.2974.20
GR00T-N1.530Monocular44.9844.3141.34
GR00T-N1.530Two-image45.6744.8742.85
GR00T-N1.530StereoPolicy47.1247.1246.32
GR00T-N1.5100Monocular63.5861.8460.08
GR00T-N1.5100Two-image65.2463.2361.22
GR00T-N1.5100StereoPolicy66.1766.1765.34
GR00T-N1.5300Monocular64.3060.8964.09
GR00T-N1.5300Two-image65.3163.9964.23
GR00T-N1.5300StereoPolicy67.5067.0267.37

C3. Real-world evaluation and camera setup. (rbLN, XaMT)

Our evaluation budget of 20 trials per method per task is within established practice for real-robot manipulation studies—700 rollouts across five tasks and seven methods. For example, Diffusion Policy evaluates each method on real-world Push-T over 20 episodes with matched initial conditions (Chi et al., 2024). We evaluate all methods on the same initial configurations, using one training seed per method and 20 trials per task. StereoPolicy achieves the highest observed success rate on every tabletop task. Its average success rate is 59%, compared with 45% for the strongest baselines, RGBD-3DDA and Multi-View, which should be considered as a comparably substantial and consistent improvement, especially in real-world experiments.

We hope these responses address the reviewers’ main concerns and welcome further discussion.

Area_Chair_ZitWOpenReview ↗

Reviewers-Authors Discussion Reminder

Dear Reviewers,

The review process for the paper has reached the rebuttal stage, and the authors have submitted their responses. Reviewers are encouraged to engage actively with these responses. Please clarify whether your initial concerns have been addressed and indicate if any new questions have arisen. Your meaningful participation is essential to maintaining a healthy ecosystem for ICLR.

Your AC.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.