acceptodds
Under review as a conference paper at ICLR 2027

BREE: A Low-Cost, Fully Fabricable and Reproducible Platform for Real-World Bimanual Robotic Manipulation

Abstract

Scaling data has been a key driver of recent advances in embodied intelligence, yet real-world robotic data remain fragmented across heterogeneous embodiments and physical platform realizations. Even nominally identical hardware can produce different observation and action distributions due to variations in calibration, sensing, workspace geometry, and task environments. This makes independently collected data difficult to aggregate and reuse. Here we introduce BREE, a low-cost () and fully reproducible bimanual robotic platform built on dual SO-101 arms. BREE standardizes the embodiment, sensing configuration, workspace, and calibration process, supported by BREE-Assets and BREE-Data. BREE-Assets provides standardized and fully fabricable task objects, fixtures, environmental components, including a dedicated physical calibrator for cross-instance alignment. BREE-Data provides eight standardized bimanual manipulation tasks with demonstrations and evaluation protocols based on BREE-Assets. Extensive cross-instance evaluations demonstrate that policies trained on data from one BREE instance can be directly transferred to another, while aggregating data collected from multiple instances can improve policy performance. That is, BREE exhibits cross-instance transfer and multi-instance scaling properties, providing a practical path toward scalable real-world robot learning across distributed platforms. We will open-source the complete BREE stack.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Real-World Robot Data, Reproducible Robotics, Bimanual Manipulation, Robot Benchmarking

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

BREE introduces a roughly $624 bimanual platform with standardized hardware, calibration, printable task assets, and eight manipulation tasks. Experiments evaluate five policies and investigate policy transfer and demonstration pooling across two independently assembled instances.

Critical strengths

Accessible and comprehensive infrastructure. BREE standardizes the complete experimental setup, including sensing, lighting, calibration, and task assets. This addresses practical barriers to reproducing real-world manipulation experiments at low cost.

Useful benchmark design. Eight tasks cover different coordination patterns, with explicit evaluation protocols and stage-level scoring. Evaluating five policy families demonstrates the benchmark’s ability to distinguish capabilities and diagnose failures.

Critical weaknesses

Cross-instance reproducibility is only partially demonstrated. Evaluation involves two instances, and substantial performance differences remain: SmolVLA’s Plate Loading success falls from 95% to 65%, while RDT’s Cylinder Insertion success falls from 40% to 20% (Table 8). Preserved intermediate-stage performance does not establish reproducibility of complete-task outcomes. The four inference seeds also do not measure variability across independently trained policies. Evidence for multi-instance scaling is insufficient. Pooling experiments primarily concern one task and do not demonstrate a consistent advantage at matched data budgets. In Table 5, 25 demonstrations from each instance achieve 55% on instance A, versus 65% using 50 demonstrations from A alone. Adding five target-instance demonstrations on another task demonstrates useful adaptation, but does not isolate a general scaling benefit.

Questions for authors

What are the quantitative calibration errors and their variability across repeated assembly/calibration? Were the two instances built by independent operators, and how variable are cross-instance performance differences across independent training seeds?

4 comments
AuthorsOpenReview ↗

Response to Reviewer FyjT

We thank the reviewer for the careful assessment and constructive feedback on cross-instance reproducibility and multi-instance scaling. We have conducted additional experiments to examine cross-instance performance under different training seeds, multi-instance data aggregation, and geometric consistency after calibration. We address the concerns below.

W1: Limited evidence for cross-instance reproducibility and differences in full-task success rates.

Response: BREE enables cross-instance policy transfer, with several policies maintaining similar full-task success rates across independently assembled instances. As shown in Table 8, ACT and π0.5 achieve identical success rates on Switch Pushing across both instances, while π0.5 maintains 95% success on Instance B compared with 100% on Instance A for Cylinder Insertion.

Table 8. Benchmark reproduction across BREE instances. The same policy checkpoints are evaluated on Instances A and B, reporting average stage success / full-task success (%).

CoordinationTaskInstanceACTRDTSmolVLAX-VLAπ₀.₅
SynchronizedSwitch PushingA70.0/70.045.0/45.080.0/80.070.0/70.075.0/75.0
B70.0/70.040.0/40.075.0/75.045.0/45.075.0/75.0
AsymmetricPlate LoadingA65.0/30.067.5/50.097.5/95.062.5/30.092.5/85.0
B60.0/20.062.5/35.082.5/65.062.5/25.085.0/70.0
SequentialCylinder InsertionA66.7/15.068.3/40.098.3/95.048.3/10.0100.0/100.0
B66.7/5.065.0/20.086.7/80.043.3/5.098.3/95.0

For relatively long-horizon bimanual manipulation, full-task success alone cannot fully reflect task execution progress. We therefore use stage-wise success as a complementary metric to capture completed operations and identify where failures occur. As discussed in Section 4.4, stage-wise success characterizes the manipulation capabilities retained after cross-instance transfer, while full-task success measures complete-task outcomes. Together, they provide a more comprehensive evaluation of cross-instance reproducibility.

W2: Variability Across Training Seeds inference seeds do not measure variability across independently trained policies

Response: We conducted additional experiments using four different training seeds (1000–1003) for SmolVLA on Plate Loading, with all policies trained on demonstrations collected from Instance A. Each trained policy was evaluated on both BREE instances using the same four inference seeds (0–3) as in the original evaluation, with 20 rollouts per instance. The results show that cross-instance performance remains consistent across different training seeds.

Table R1. Full-task success rates (%) across training seeds (1000-1003) using 50 demonstrations from Instance A.

Training SeedInstance AInstance BA–B Gap (pp)
1000956530
1001906525
1002704525
1003855530
Mean85.057.527.5

Although success rates vary across training seeds, the cross-instance gaps remain stable at 25–30 percentage points, with an average gap of 27.5 percentage points. This indicates consistent cross-instance performance across independently trained policies, despite training variability.

W3. Evidence for data scaling is insufficient.

Response: As suggested, we have included more evidence to demonstrate BREE's data scaling (see Table R2). We included ACT on Pick-and-Place, SmolVLA on Drawer Placement, and X-VLA on Trajectory Pushing. Training budgets range from 25 to 100 demonstrations, with all evaluations conducted on Instance A.

The following results have further proved the multi-instance scaling property in our framework, i.e., demonstrations collected independently on different BREE instances can be directly combined for policy training, while maintaining performance close to training with data from a single instance.

As indicated in Table R2, when the mixed-instance dataset increases from 50 to 100 demonstrations, success rates improve from 60% to 80% for ACT, 45% to 75% for SmolVLA, and 45% to 60% for X-VLA.

Table R2. Full-task success rates (%) evaluated on BREE Instance A under different training data configurations. A and B denote two independently assembled BREE instances.

Total DemosTraining DataACT Pick-and-PlaceSmolVLA DrawerX-VLA Trajectory
25A25453025
50A50604550
A25+B25604545
75A75806555
A25+B50806555
100A100858070
A25+B75807560
AuthorsOpenReview ↗

Response to Reviewer FyjT

Q1: Lack of quantitative calibration error and repeatability measurements.

Response: We conducted two additional experiments to quantitatively evaluate BREE's calibration accuracy and repeatability: cross-instance visual geometric consistency across workspace configurations and calibration repeatability under repeated robot-arm reassembly.

Cross-instance visual alignment achieves an overall error of 18.20 mm-equivalent RMS. We evaluated two independently assembled BREE instances, each with two SO-101 arms, at three predefined joint configurations (P1–P3), with three repeated executions per configuration (36 measurements in total). AprilTag image-center displacements were measured relative to Instance A references and converted to millimeter-equivalent distances using the reference AprilTag scale.

Table R3. Visual geometric consistency (mm-equivalent).

MetricLeft ArmRight ArmOverall
BREE: Cross-instance RMS14.7321.1118.20
BREE: Maximum cross-instance deviation18.1628.05—
Standard SO-101: Cross-instance RMS——65.77

BREE calibration achieved an overall cross-instance RMS error of only 18.20 mm-equivalent, compared with 65.77 mm-equivalent under standard SO-101 calibration, representing a 72.3% reduction. Despite combined geometric variations from the robot arms, camera mounts, and aluminum frames, BREE maintains centimeter-level cross-instance visual alignment.

As shown in Table 3 of the paper, BREE calibration achieves 90% success after arm replacement, compared with 0% using standard SO-101 calibration, demonstrating its effectiveness in maintaining policy performance across robot arms.

Repeated reassembly and recalibration produce consistent visual alignment for the same arm. We evaluated three SO-101 arms (Arms 1, 2, and 3), each undergoing three independent cycles of disassembly, reinstallation, and BREE recalibration. After each cycle, AprilTag image-center positions were measured at the same three fixed locations, yielding 27 alignment records.

Table R4. Within-arm visual displacement across repeated reassembly and recalibration cycles (px).

MetricArm 1Arm 2Arm 3Overall
Mean within-arm displacement5.95.77.86.5
Maximum within-arm displacement8.610.715.815.8

The mean within-arm displacement across repeated reassembly and recalibration was 6.5 px, with a maximum of 15.8 px. These results demonstrate consistent visual alignment after independent reassembly and recalibration.

Q2: Independent Operators

Response: BREE Instances A and B were independently assembled and calibrated by two different operators, following the same construction specifications and calibration procedures. The cross-instance evaluation therefore includes differences introduced by independent assembly and calibration, although the experiments are limited to two physical instances.

Dear Authors,

Thanks for your comment and the added experiments on the cross-instance validation and supplement evidence. Most of my concerns are resolved, and I've raise my rating to a weak acceptance.

AuthorsOpenReview ↗

Thank you for your support

Thank you for your positive feedback and for recognizing our work. Your constructive comments have helped us improve our experiments and strengthen the paper. We truly appreciate your time and suggestions!

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

This paper introduces BREE, an open source hardware platform for low-cost, reproducible real-world bimanual robot manipulation and evaluation. The hardware system uses two SO-101 arms, as well as a full environment set up that is inexpensive and relatively easy to assemble, with minimal unique parts. They include tools to aid calibration, and methods to vary the environment, like initial configuration references, LED lights to vary lighting conditions, and calibration methods to maintain consistency across instances. The hardware system is analyzed on two separately assembled instances on five modern imitation learning policies to show policy transfer across instances.

Critical strengths

The fully open sourced hardware system that is low cost, accessible, and includes tools to help with repeatability for both the environment and the assets is the strength of this paper. Their problem statement that robotics lacks a lot of standardized, real-world evaluation to compare different policies is very well motivated, and their hardware environment is in the right direction to support that effort. They also have some unique features that help to evaluate the robustness of different models, such as LED lights that can test robustness to lighting conditions in a controlled manner.

Critical weaknesses
  1. Standardized evaluation protocol includes a set of reference scene images to reproduce configurations across BREE instances, which is helpful and interesting for having reproducible and consistent real world evaluations that allow methods to compare against each other fairly. However, these are only given references, and it would have been nice if there was some verification process or video submission to a 3rd party oracle that will check if the setup is actually close to the given initial configuration. Without this, there is no guarantee that the physical evaluations are actually run properly.

  2. The inexpensive hardware offers accessibility, but somewhat against the motivation of reproducibility. These servo motors do not have precision or accuracy. The servo specification sheets report around 0.5deg of backlash, with 3rd party cases report more. 3D printed tolerances also have some variability, which would add to the variance in reproducibility. A study on the accuracy, precision, and repeatability of the SO-101 arms across instances would have been needed for verification. Using widely adopted hardware, such as the bimanual YAMs, and applying a similar environment setup and task assets could be better for real-world reproducibility and evaluations.

Questions for authors
  1. The reference scene images are helpful for reproducing the initial configurations across BREE instances. However, how do you check if the setup is actually close enough to the given reference? Is there any verification process or tolerance for object placement, camera position, and arm configuration to ensure that different groups are running comparable evaluations?
  2. The budget hardware is great for accessibility, but I wonder how much the servo backlash and 3D-printed tolerances affect reproducibility. Do you have measurements of the accuracy, precision, and repeatability of the SO-101 arms across the two instances? It would be helpful to understand how much of this variability the calibration methods can actually address.
  3. How does the variation in policy performance across the two instances compare to the variation from running the same policy multiple times on one instance? How many evaluation trials were run, and do the five policies have consistent relative performance across instances? This would help show whether the differences are coming from the policies or from the hardware and setup.
  4. Were the two instances independently assembled and calibrated using only the provided instructions and tools? For policy transfer across instances, were any additional manual adjustments, demonstrations, or fine-tuning needed? I wonder how much effort and prior knowledge another group would need to reproduce the same setup and results.
3 comments
AuthorsOpenReview ↗

Response to Reviewer edSV

We sincerely thank the reviewer for recognizing BREE's open-source, low-cost design and its value for reproducible real-world evaluation. We greatly appreciate the constructive feedback and address each concern below.

W1 & Q1:Lack of objective verification and quantitative tolerances for reproducing evaluation setups across BREE instances.

Response: As suggested, we have added a clear verification process to the BREE assembly and calibration checklist, with empirically informed acceptance criteria to ensure consistent setup verification across independently assembled instances.

What we verifyWhenHow we verify itSuggested Acceptance Criteria
Workspace and camera alignmentAfter assemblyCompare fixed landmarks on the frame and robot mounts in the current global camera image with the reference image.Mean image displacement < 1 px. The measured displacement was 0.577 px in a representative test.
Robot home poseAfter calibrationCompare the home pose with the physical reference established by the BREE Calibration Aid.Height deviation ≤ 3 mm and upper arm inclination deviation ≤ 5°.
Task object placementBefore each rolloutUse the designated fixtures where available and compare object position and orientation with the reference image.Objects are seated in their fixtures where applicable. For freely placed objects, operators correct visible deviations. Numerical placement tolerances are currently unspecified.

The proposed verification process allows independent teams to check workspace geometry, robot configurations, and task initialization against the same documented criteria before evaluation.

We have updated the checklist and the object placement procedure in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer edSV

W2 & Q2:Lack of quantitative validation of SO-101 accuracy and repeatability.

Response: As suggested, we conducted two additional experiments to quantify geometric repeatability and cross-arm consistency, and compared BREE calibration with the standard SO-101 procedure.

1. BREE calibration achieves repeatable configurations across two independently assembled instances.

We evaluated both arms on Instances A and B at three predefined joint configurations (P1–P3), with three repetitions per configuration, yielding 36 measurements. Using 40 mm AprilTags observed by the wrist cameras, we measured image-center deviations and converted them to millimeter-equivalent distances using the reference AprilTag scale. RMS (root mean square) error summarizes the AprilTag center displacements relative to the reference observations.

Across the three configurations, BREE calibration achieved an overall cross-instance RMS error of only 18.20 mm-equivalent, compared with 65.77 mm-equivalent under standard SO-101 calibration, representing a 72.3% reduction.

Despite combined geometric variations from the robot arms, camera mounts, and aluminum frames, BREE maintains centimeter-level cross-instance visual alignment.

Table R1: Visual geometric consistency after BREE calibration (mm-equivalent).

MetricLeft ArmRight ArmOverall
BREE: Cross-instance RMS14.7321.1118.20
BREE: Maximum cross-instance deviation18.1628.05—
Standard SO-101: Cross-instance RMS——65.77

2. BREE calibration improves reproducibility after arm replacement and recalibration.

We further tested three independently fabricated SO-101 arms after reinstallation and recalibration, using three predefined configurations and three repetitions per arm (27 trials).

Under standard SO-101 calibration, only 7/27 measurements (25.9%) produced valid AprilTag observations, compared with 27/27 (100%) under BREE calibration. No arm under standard calibration yielded valid observations at all three predefined configurations.

Table R2. Visual repeatability after arm replacement and recalibration.

MetricStandard SO-101BREE Calibration
Valid AprilTag observations7/27 (25.9%)27/27 (100%)

The policy results in Table 3 further demonstrate this improvement. ACT achieved 90% success after arm replacement with BREE calibration, compared with 0% under standard SO-101 calibration (10 trials per condition).

Our BREE Calibration Aid establishes a shared physical reference, while predefined joint ranges and instance-specific homing offsets align joint-coordinate mappings across arms. Although this procedure cannot eliminate servo backlash or fabrication tolerances, it improves configuration reproducibility and enables more reliable cross-arm policy transfer despite these mechanical differences.

3. Cost considerations for alternative hardware.

We appreciate the reviewer's suggestion to consider more precise platforms such as YAMs. While a single YAM arm costs approximately $3,000, the complete bimanual BREE platform costs approximately $600. Adopting higher-cost hardware could improve mechanical precision but would also increase the cost of reproducing the platform, which is an important consideration for accessibility.

AuthorsOpenReview ↗

Response to Reviewer edSV

Q3: Lack of quantitative comparison of within-instance variability and cross-instance policy performance

Response: Our evaluations show that within-instance performance fluctuations are typically around 5 percentage points, compared with a mean absolute cross-instance difference of 11.0 percentage points. The five policies also largely preserve their relative performance across BREE instances.

3.1. Cross-instance performance variation is larger than the typically observed within-instance fluctuation.

Repeated evaluations of the same policy on the same instance typically showed success-rate fluctuations of approximately 5 percentage points. In comparison, Table 8 reports a mean absolute cross-instance difference of 11.0 percentage points across 15 policy–task combinations, with task-level averages of 7.0, 15.0, and 11.0 percentage points for Switch Pushing, Plate Loading, and Cylinder Insertion, respectively.

As discussed in Q2, residual geometric differences remain between independently assembled robot instances despite BREE calibration. These hardware variations, together with differences in policy robustness, may contribute to the observed cross-instance performance differences.

3.2. All five policies are evaluated under matched conditions with 20 rollouts per instance.

As described in Sec. 4.1 and Appendix B.2, each policy is evaluated on five predefined scene configurations with four fixed inference seeds (0–3), yielding 20 rollouts per policy per instance. Both instances use the same configurations, seeds, checkpoints, and task-specific execution budgets, without additional training or fine-tuning.

3.3. The five policies show consistent overall performance trends across instances.

As shown in Table 8, the five policies exhibit consistent overall performance trends across Instances A and B. The relative performance patterns are largely preserved across all three tasks, with stronger-performing policies generally maintaining their advantages on both instances. These results support BREE's ability to provide consistent comparative evaluations across independently assembled robot instances.

Q4:Independent Operators

The two BREE instances were independently assembled and calibrated by different individuals using the same documented procedures, and direct cross-instance policy transfer required no additional demonstrations or fine-tuning.

1. Both BREE instances were independently assembled and calibrated by different individuals.

We provide detailed dimensional drawings, component specifications, and step-by-step assembly guides covering the frame geometry, robot mounting positions, camera placement, and workspace layout. Both instances were calibrated using our BREE Calibration Aid and standardized procedures, with a reproduction checklist provided in Appendix A.3.

2. Direct cross-instance transfer required no additional demonstrations, retraining, or policy adjustments.

Policies trained on Instance A were directly deployed on Instance B using the same checkpoints, without target-instance demonstrations or fine-tuning. Only the standard calibration, camera alignment, and task initialization procedures were performed.

Furthermore, as shown in Table 3, incorporating just five demonstrations from Instance B improved the full-task success rate of Trajectory Pushing from 0% to 35%, with average stage success increasing from 42% to 60%. This demonstrates that a small amount of target-instance data can further improve policy performance beyond direct cross-instance transfer.

3. Reproducing BREE requires only basic assembly and calibration skills.

The setup involves mechanical assembly, 3D printing, robot calibration, and camera alignment. Our dimensional drawings, assembly guides, calibration tools, and reproduction checklist document these steps, reducing reliance on prior experience or undocumented adjustments.

We have clarified these details in the revised manuscript.

Reviewer_z6nSOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces BREE, a low-cost bimanual platform built on dual SO-101 arms. It standardizes robot calibration, cameras, lighting, and fabricable task assets to reduce differences between independently assembled setups. The authors establish an eight-task benchmark, evaluate five policies, and investigate policy transfer and data aggregation across two physical instances on selected tasks.

Critical strengths
  1. Practical value. The low-cost hardware, calibration fixtures, task assets, and unified evaluation protocol can lower the barrier to controlled real-world robot learning experiments.

  2. A reasonably complete system. The platform integrates the main components needed for reproducible experiments. Hardware replacement and cross-instance evaluations provide preliminary evidence that policies and data can be reused across homogeneous setups.

Critical weaknesses
  1. The contribution is predominantly engineering integration. The work primarily integrates existing hardware, calibration procedures, and evaluation tasks. It does not introduce a novel method or establish new knowledge or generalizable research findings beyond existing understanding.

  2. Evidence for data scaling is insufficient. The data aggregation experiments involve only two instances and a small number of tasks, primarily using ACT. They do not demonstrate consistent benefits as the number of instances or the dataset size increases. The results support the feasibility of pooling data, but lack systematic scaling curves, sufficiently comprehensive comparisons at matched data budgets, and statistical validation.

  3. Cross-instance reproducibility mainly demonstrates an engineering benefit. Policy reuse after standardizing hardware, calibration, and the environment demonstrates the engineering effectiveness of the setup. However, the paper does not derive a new method or generalizable scientific insight from these results. In its present form, cross-instance reproducibility therefore does not provide a sufficient basis for the central research contribution.

Questions for authors
  1. Please clarify the novelty of this work: what new knowledge or research insights does it provide to the community beyond prior work?
5 comments
AuthorsOpenReview ↗

Response to Reviewer z6nS

We thank the reviewer for the thoughtful evaluation and constructive feedback on the novelty and scientific contributions of our work. We address the concerns below.

W1 & Q1 & W3: The contribution is predominantly engineering integration, and It does not introduce a novel method.

Response: We would like to clarify that BREE's primary research contributions lies in two folds: 1) addressing a largely unexplored challenge: facillatating data collected from independently constructed robotic platforms to be directly reused and aggregated for policy learning. 2) community contribution: fostering broader community participation in embodied intelligence by providing an affordable, reproducible, and accessible research platform, with resuable data from indivisual instances across the world. These contributions do NOT necessarily come from a particular new algorithm or policy.

Unlike existing efforts that primarily focus on reproducible evaluation, BREE addresses the largely unexplored challenge of cross-instance training-data compatibility. Hence, we introduce BREE-Assets, a standardized and fully fabricable collection of physical components. These physical components, including but not limited to new-designed physical calibrator, unified structural frame and camera configuration, and fully fabricable task objects and fixtures, are specifically designed to minimize cross-instance discrepancies in joint-coordinate mappings, sensing configurations, workspace geometry, and task environments. All these components are integrated organically in a sophisticated manner.

Beyond its technical contribution, BREE lowers the barrier to real-world embodied intelligence research by providing an affordable (~$600), fully reproducible, and open-source bimanual robotic platform. More importantly, it enables researchers worldwide to independently construct compatible platforms and contribute reusable demonstration data, laying the foundation for collaborative, large-scale data collection and community-driven robot learning.

W2: Evidence for data scaling is insufficient.

Response: As suggested, we have included more evidence to demonstrate BREE's data scaling (see Table R2). We included ACT on Pick-and-Place, SmolVLA on Drawer Placement, and X-VLA on Trajectory Pushing. Training budgets range from 25 to 100 demonstrations, with all evaluations conducted on Instance A.

The following results have further proved the multi-instance scaling property in our framework, i.e., demonstrations collected independently on different BREE instances can be directly combined for policy training, while maintaining performance close to training with data from a single instance.

As indicated in Table R2, when the mixed-instance dataset increases from 50 to 100 demonstrations, success rates improve from 60% to 80% for ACT, 45% to 75% for SmolVLA, and 45% to 60% for X-VLA.

Table R2. Full-task success rates (%) evaluated on BREE Instance A under different training data configurations. A and B denote two independently assembled BREE instances.

Total DemosTraining DataACT Pick-and-PlaceSmolVLA DrawerX-VLA Trajectory
25A25453025
50A50604550
A25+B25604545
75A75806555
A25+B50806555
100A100858070
A25+B75807560

Thank you for taking the time to run these additional experiments. The results across more policies and data budgets give better support for pooling data from the two instances. I appreciate the work that went into building and testing BREE, and I see its low cost and reproducibility as useful contributions to the robotics community.

My concerns about novelty and scaling remain, though. A new algorithm is not necessary, but I am still unsure what new technical idea or research insight the paper establishes beyond the standardized setup. The scaling experiments show that adding demonstrations from a second instance improves performance. This is useful evidence, but with only two fixed instances, it does not yet support the paper’s broader claims about scaling across distributed platforms.

I will keep my rating. The additional experiments have addressed part of my concern, but they have not changed my assessment of the paper’s research contribution.

AuthorsOpenReview ↗

Further clarificaiton regarding research insights

We thank the reviewer for clarifying the remaining concern regarding the research insights and technical idea beyond engineering standardization. As suggested, we are able to summarize the research insights in this paper as follows (some of which are found owing to the constructive comments provided by Reviewer edSV and Reviewer FyjT!).

We highlight three findings the BREE offers below.

1. Effective Cross-Instance Policy Transfer Despite Mechanical Discrepancies.

Our results show that effective policy transfer does not require mechanically identical robot instances. Despite fabrication tolerances and assembly variations, BREE calibration reduces cross-instance visual geometric error from 65.77 to 18.20 mm-equivalent RMS, a 72.3% reduction. More importantly, ACT achieves 90% success after arm replacement, compared with 0% under standard SO-101 calibration.

Table R3: Calibrations in BREE

Metric (mm-equivalent)Left ArmRight ArmOverall
BREE: Cross-instance RMS14.7321.1118.20
BREE: Maximum cross-instance deviation18.1628.05—
Standard SO-101: Cross-instance RMS——65.77

2. Fine-Grained Manipulation Is More Sensitive to Cross-Instance Discrepancies Than Basic Manipulation.

Cross-instance transfer does not affect all manipulation capabilities equally. As shown below, policies retain substantial stage-level performance even when full-task success declines. This suggests that basic manipulation behaviors are more transferable, whereas fine-grained execution is more vulnerable to residual physical discrepancies.

Table R4: Stage-wise Transfer v.s. Full-Task Sucess

Policy / TaskStage Success (A → B)Full-Task Success (A → B)
ACT / Trajectory Pushing55% → 42%15% → 0%
RDT / Cylinder Insertion68.3% → 65.0%40% → 20%

3. Cross-Instance Demonstrations Improve Policy Learning Despite Hardware Discrepancies.

Demonstrations collected from independently assembled instances remain useful for policy training despite residual hardware differences. With 25 demonstrations from Instance A fixed, adding 75 demonstrations from Instance B improves the success rates of SmolVLA from 30% to 75% and X-VLA from 25% to 60%.

Table R5: Multi-instance Scaling Demonstrations

Total DemosTraining DataSmolVLA DrawerX-VLA Trajectory
25A253025
50A25+B254545
75A25+B506555
100A25+B757560

These findings suggest that cross-instance policy transfer and data reuse are feasible despite residual mechanical discrepancies. We acknowledge that further validation across more instances is needed. Nevertheless, we have repeatedly disassembled, reassembled, and recalibrated utilizing these two existing instances. This could serve partially as evidence for for reproducibility across independent reconstruction cycles. A third instance is currently under construction, though its evaluation may not be completed within the rebuttal period.

Please let us know if our response this time addresses your concern. We genuinely appreciate your constructive feedback, which has helped us move beyond our original benchmark-focused scope and substantially strengthen the paper with deeper research insights.

Comments on the paper

1 comment
Area_Chair_EAt9OpenReview ↗

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.