RAPID Hand: Robust, Affordable, Perception-Integrated, Dexterous Manipulation Platfrom for Embodied Intelligence

We thank the reviewer for the detailed reading of our paper and constructive suggestions! We hope our responses adequately address the following questions about our work. Please let us know if there’s anything we can clarify further.

1. As RAPID Hand requires manual assembly, even if the bill of materials is low, some hidden costs, such as time, tooling (e.g., 3D printer), and expertise required for assembly are non-trivial and may limit accessibility for some labs. In terms of labor, time, and expertise, what does it take to assemble a RAPID Hand?

Reply: Manual assembly inevitably involves non-trivial costs in time, tooling, and expertise beyond the bill of materials, a challenge shared by other hardware platforms such as LEAP Hand (about 4 hours to assemble [1]). In our experience, assembling a RAPID Hand takes about 6 hours for an undergraduate student familiar with standard prototyping tools. Although RAPID Hand includes slightly more components than LEAP Hand, the overall assembly difficulty is comparable. The required equipment is limited to a desktop 3D printer and common tools such as soldering kits and screwdrivers.

To lower the barrier, we design the hand with modular components and provide detailed open-source CAD files, wiring diagrams, and assembly guides. While some effort is unavoidable, we believe the overall accessibility remains significantly higher than that of existing dexterous hands. To further lower the entry barrier, we plan to release step-by-step assembly instructions and partially assembled kits in future updates.

2. The paper does not mention repeatability by replaying collected trajectories, which should be a straightforward evaluation of the accuracy of the entire system. Can the system fully replay a recorded trajectory?

Reply: Yes, the system can fully replay a recorded trajectory. Repeatability involves both the arm and the hand. For the arm, we use the widely adopted UR10e, which offers a pose repeatability of ±0.05 mm [2]. For the hand, we reported joint precision during motion (Fig. 8) and performance under varying loads (Fig. 9) in the appendix, both demonstrating consistent behavior. Moreover, RAPID Hand has autonomously executed repeated public demonstrations, further supporting the system’s repeatability.

3. While the paper emphasizes the utility of wrist-mounted vision for in-hand perception, it does not discuss the compatibility of RAPID Hand with diverse robot arms or full-body humanoid platforms. As dexterous hands are increasingly deployed in integrated systems, the lack of evaluation or guidance on cross-platform adaptability may limit the system's practical adoption. Can the hand be equipped on mainstream robot arms and bodies? In the paper, the hand is equipped on UR10e. But it does not discuss integration with other arms or humanoid platforms, nor does it provide mounting specs or interface standards.

Reply: RAPID Hand can be mounted on other robot arms and humanoid platforms, though some additional adapters are required. The hand uses a modular wrist structure in which the camera mount remains fixed, and only the wrist flange needs to be replaced when adapting to a different arm. Both the flange and the camera bracket follow standard mechanical interfaces, which facilitates integration across platforms.

We are preparing wrist flanges for mainstream collaborative arms (e.g., Franka, xArm) and humanoid platforms (e.g., Unitree G1), together with detailed mounting specifications and interface standards to further support cross-platform use.

4. How robust is the hand? How often does it break or burn out? How does the system handle sensor degradation or calibration drift over time? The paper shows stable tracking under load, no overheating, and larger load capacity over LEAP Hand. However, failure rates or long-term durability (e.g., motor burnout, sensor degradation) are not reported.

Reply: Based on our experience, RAPID Hand has remained generally robust under daily use for about six months. The only notable incident occurred when the arm lost control and the hand struck the table, breaking a finger linkage. The part was replaced within about 5 minutes, and the system continued to operate normally. In tasks such as multi-finger retrieval, where fingers frequently interact with stacked objects in confined spaces, the hand has so far functioned reliably without structural failures.

We have not observed clear signs of degradation in the tactile sensors to date. For proprioception, we use a custom calibration board to reset each phalange, typically once per day, to maintain accuracy. Current-based position control also helps reduce the risk of motor burnout from stall-induced overheating. We will include these details in the paper to provide a clearer picture of the system’s robustness and durability.

5. How transferable is the data or policies on other widely used dexterous hands (e.g., Allegro, Shadow) to RAPID Hand?

Reply: Cross-hand transfer remains a significant challenge in the field. Differences in actuation and transmission mechanisms, hand size, and sensing modalities (e.g., optical, piezoresistive, capacitive tactile sensors) all contribute to distributional gaps that affect transferability. For example, the Allegro hand lacks tactile sensing, while the Shadow hand is tendon-driven, making direct policy transfer difficult.

A practical direction we see is to use data from other hands (e.g., Allegro, Shadow) as part of training robot foundation models [3, 4, 5], and then fine-tune with a smaller set of RAPID Hand demonstrations. While we have not yet conducted such experiments, we consider this an important avenue for future work.

[1] Shaw, Kenneth, Ananye Agarwal, and Deepak Pathak. "Leap hand: Low-cost, efficient, and anthropomorphic hand for robot learning." Robotics: Science and Systems (RSS 2023).

[2] https://www.universal-robots.com/media/1807466/ur10e_e-series_datasheets_web.pdf

[3] Physical Intelligence, et al. “π0.5: a vision-language-action model with open-world generalization.” arXiv preprint arXiv:2504.16054, 2025.

[4] Gemini Robotics Team, et al. “Gemini robotics: Bringing ai into the physical world.” arXiv preprint arXiv:2503.20020, 2025.

[5] Johan Bjorck, et al. “GR00T-N1: An open foundation model for generalist humanoid robots.” arXiv preprint arXiv:2503.14734, 2025.

We thank the reviewer for the detailed reading of our paper and constructive suggestions! We hope our responses adequately address the following questions about our work. Please let us know if there’s anything we can clarify further.

1. Limited task complexity: While the three evaluated tasks demonstrate basic dexterous capabilities, they are relatively simple compared to real-world manipulation needs. Tasks like tool use, bimanual coordination, or fine assembly would better demonstrate the platform's capabilities for "generalist robot autonomy."

Reply: The three evaluated tasks, namely in-hand rolling, translation, and multi-finger retrieval, are chosen as atomic skills that test durability, sensing consistency, and dexterous control. More complex tasks such as tool use can be viewed as compositions of these fundamentals. RAPID Hand is designed as an affordable, reliable platform for consistent multimodal data collection. While advancing to more complex tasks would further demonstrate its potential, progress is constrained by the difficulty of collecting safe, high-quality demonstrations, since human teleoperation without tactile feedback makes producing dexterous motions on a different embodiment particularly challenging [1]. As a next step, we are exploring RAPID Glove, a high-Resolution, Affordable, haPtic Interface for Demonstrating contact-rich manipulation, as an important direction for extending the platform’s capabilities.

2. Perception limitations: The flat tactile sensors provide only normal force measurements, lacking shear force sensing crucial for slip detection and fine manipulation. The single wrist-mounted camera may suffer from occlusions during complex manipulations.

Reply: The current system uses flat fingertip tactile arrays to balance affordability with consistent performance. While shear force sensing can improve slip detection and fine manipulation, our tests with sensors from Paxini [2], TashanTec [3], and Xela Robotics [4] confirmed such benefits but also showed significant increases in cost, maintenance effort, and difficulty in ensuring high-resolution, consistent readings. Given our focus on cost-effectiveness and robustness, we chose flat arrays for this design.

For vision, we recognize that a single wrist-mounted camera may lead to occlusions. Our perception framework can readily support hardware-synchronized external cameras, and we plan to explore this extension in future work to improve coverage during complex manipulations.

3. Teleoperation latency: What is the end-to-end latency from human hand motion to robot execution? How does the optimization solver perform under rapid movements or when approaching joint limits?

Reply: The end-to-end latency from human hand motion to robot execution is mainly due to the UR collaborative arm, typically around 1–2 s. The RAPID Hand itself responds with minimal delay. While this latency is noticeable, teleoperators generally adapt after some practice, allowing us to collect demonstrations reliably.

For rapid movements, the optimization solver can compute inverse kinematics solutions, but execution remains limited by the arm’s delay. To reduce risks of collision or hardware strain, we avoid aggressive motions and prevent the arm from approaching joint limits. We acknowledge that reducing arm latency would further improve teleoperation, and this could be addressed by adopting a robotic arm with faster execution in future work.

4. Tactile sensor durability: Given that flat piezoresistive sensors are "inevitable to wear and tear," what is the expected lifetime under typical manipulation loads? How does sensor degradation affect policy performance?

Reply: Given that piezoresistive sensors are inevitably subject to wear and tear, we believe their durability is largely determined by fabrication processes and quality. For reference, we previously used the tactile sensor in [5], which maintained stable performance for about three months of daily manipulation before gradually declining. The sensors used in this work, according to supplier testing, can withstand over one million impacts at a 10 kg load. Based on our experience, since their deployment in November 2024, we have not needed to replace any, and no noticeable degradation has been observed.

If replacement is required, each sensor is mass-produced and costs about $70** [6], in contrast to alternatives such as Paxini [2] (**$400) and Xela USkin [4] ($2,000), which helps keep maintenance affordable. To date, RAPID Hand has consistently performed in-hand translation and rolling tasks, including repeated public demonstrations, without evident loss of performance.

5. Force control capabilities: The system uses position control throughout. How does this limit performance in tasks requiring precise force regulation, such as fragile object handling or assembly?

Reply: Using position control alone has inherent limitations, especially for tasks that require precise force regulation. In our evaluations, we occasionally observed excessive forces, such as minor surface damage to vegetables during in-hand rolling. In this work, we adopt position control to validate the stability and effectiveness of the RAPID Hand platform for autonomous policy learning, as it allows reliable and consistent execution across a range of contact-rich tasks, though we note that force control will be essential for more delicate tasks such as fragile object handling or fine assembly and plan to pursue this in future work.

[1] Yin, Zhao-Heng, et al. "DexterityGen: Foundation controller for unprecedented dexterity." Robotics: Science and Systems (RSS 2025).

[2] https://paxini.com/

[3] https://www.tashantec.com/

[4] https://www.xelarobotics.com/tactile-sensors

[5] Lu, Peng, et al. "Thermoformed electronic skins for conformal tactile sensor arrays." 2024 IEEE International Conference on Robotics and Automation (ICRA 2024).

[6] https://www.moxiantech.com/robot-scene/14

We thank the reviewer for the detailed reading of our paper and constructive suggestions! We hope our responses adequately address the following questions raised about our work. Please let us know if there’s anything we can clarify further.

1. The paper emphasizes the characterization of the hardware system but devotes much less attention to the unique capabilities and advantages that the new hardware offers. This imbalance may reduce its appeal to the broader NeurIPS audience. 

Reply: We conclude the unique capabilities of RAPID Hand as follows.

First, the hardware is co-designed with perception and teleoperation to address practical challenges in dexterous manipulation data collection. The universal multi-phalangeal actuation scheme with a bevel gear differential achieves a compact 20 mm finger thickness while maintaining robustness, and the parallel MCP joint provides 2.3× higher load capacity than the LEAP Hand’s tandem design.

Second, on the sensing side, hardware-level synchronization ensures sub-7 ms multimodal latency, and fingertip tactile signals are spatially aligned into local “touch point clouds,” enabling direct geometric correspondence not available in prior tactile systems.

Third, these capabilities translate into practical benefits: in in-hand translation, RAPID Hand achieves smooth lateral finger motion, while Allegro often drops the object and LEAP shows minimal movement. Compared with prior work [1], our setup also removes simplified assumptions such as fixed end-effectors and table support, as shown in the supplementary video.

We will devote more attention to highlighting these unique capabilities in the final version.

2. Certain critical aspects, such as system latency, are not fully validated. Particularly, in Figure 4: What is the server being used? If it is running on typical operating systems like Windows or Linux, which are not real-time systems, the reported latency values seem unconvincing.

Reply: The sub-7 ms latency reported in Figure 4 refers specifically to multimodal synchronization, not end-to-end teleoperation delay. It measures the interval between the MCU (Figure 4 “Hard Sync”) receiving fingertip haptic data and triggering the camera exposure. This was verified using a logic analyzer.

Such low latency is achieved through hardware-level synchronization (e.g., MCU-based triggering and timestamp alignment) and is independent of the upper-level operating system. While Linux is not a real-time OS, it does not affect this measurement, as the synchronization occurs entirely on the MCU. We will clarify this definition in the final version.

3. In Figure 2: Why does the paper not compare the proposed hand with the five-finger Shadow Hand? This omission appears significant. 

Reply: We compare RAPID Hand with Allegro and LEAP because they are fully actuated, direct-drive, relatively affordable, and widely adopted in recent works [3, 4, 5], making them the most relevant baselines. The Shadow Hand, while highly capable, follows a design focus different from ours, being tendon-driven and oriented toward more advanced manipulation skills. Our work emphasizes platforms suited for long-term autonomous operation and large-scale data collection, where accessibility and robustness are critical. While Shadow is not included in Figure 2, we provide a specification-level comparison in Table 2 and plan to add a size comparison in the final version.

4. In Figure 7: What does the bar represent—mean values? If so, where are the error bars or other indicators to show the distribution? Further, a significance test should be considered to support the conclusions drawn from the results.

Reply: Mean Values and Standard Deviations The bars in Figure 7 show the mean action MSE for each condition. The mean and standard deviation for each setting are summarized below:

Statistical Significance We conducted two-sample t-tests for each task, and all differences between "Ours" and "w. latency" are statistically significant (p < 0.001).

[1] Si, Zilin, et al. "Tilde: Teleoperation for dexterous in-hand manipulation learning with a deltahand." Robotics: Science and Systems (RSS 2024).

[2] Bai, Fengshuo, et al. "Retrieval dexterity: Efficient object retrieval in clutters with dexterous hand." arXiv preprint arXiv:2502.18423 (2025).

[3] Qi, Haozhi, et al. "In-hand object rotation via rapid motor adaptation." Conference on Robot Learning (CoRL 2023).

[4] Suresh, Sudharshan, et al. "NeuralFeels with neural fields: Visuotactile perception for in-hand manipulation." Science Robotics (2024)

[5] Zhang, Jialiang, et al. "Dexgraspnet 2.0: Learning generative dexterous grasping in large-scale synthetic cluttered scenes." 8th Annual Conference on Robot Learning. (CoRL 2024).

We thank the reviewer for the detailed reading of our paper and constructive suggestions! We hope our responses adequately address the following questions raised about our work. Please let us know if there’s anything we can clarify further.

1. Evaluation is mainly on a small set of tasks; broader real-world generalization needs further study. 

Reply: For autonomous manipulation with imitation learning, most existing SOTA work [1, 2, 3] focuses on pick-and-place, while our tasks, including in-hand rolling, translation, and multi-finger retrieval, involve more complex, contact-rich manipulation. We select them as atomic skills to test stability, sensing consistency, and dexterous control, providing a practical benchmark for evaluating RAPID Hand’s 20-DoF design and synchronized multimodal sensing.

For dexterous teleoperation, a concurrent SOTA work on tool use [4] relaxes several conditions, such as placing tools in grasp friendly poses, assuming the tool is already in hand, or fixing the arm for stability. Our preliminary experiments under similar settings suggest that RAPID Hand can also support such capabilities.

More complex tasks, such as full tool use or fine assembly, would further extend the platform’s potential. Progress toward these tasks remains constrained by the difficulty of collecting safe, high-quality demonstrations, particularly without tactile feedback [5]. As a natural next step, we are exploring RAPID Glove, a high-Resolution, Affordable, haPtic Interface for Demonstrating contact-rich dexterous manipulation. While this is beyond the scope of the present paper, we see it as an important direction for extending the platform’s capabilities.

2. The teleoperation interface lacks force feedback, reducing realism of remote control; have the authors evaluated how this limitation affects quality?

Reply: Tactile feedback can improve the realism and quality of teleoperation. In this work, we use a Vision Pro–based interface to ensure accessibility and ease of use. While it does not provide force feedback, it allows us to collect stable and high-quality demonstrations for all evaluated tasks.

We do not directly quantify the impact of missing force feedback, as such an evaluation would require high-resolution haptic gloves, which are currently prohibitively expensive (USD 15,000–75,000) and bulky (e.g., HaptX [6], Meta [7]). Instead, we evaluate its effect through repeated teleoperation trials and qualitative monitoring of task outcomes. Based on these observations, the absence of tactile feedback can occasionally result in excessive force, such as surface damage during rolling.

3. The tactile sensors are limited to the fingertips. Can the authors discuss how this impacts tasks involving palm contact?

Reply: In this work, we limit tactile sensing to the fingertips due to cost and integration constraints. Full-palm coverage requires custom-shaped or curved sensors, which remain costly (e.g., ReSkin [8], Paxini [9], Xela Robotics [10], Tencent [11]). Prior work shows that palm sensing can benefit certain tasks; for example, ReSkin applied to the Allegro Hand’s palm achieved nearly 100% success in in-hand translation [12].

In our case, the evaluated tasks: rolling, translation, and multi-finger retrieval, have so far remained reliable without palm sensors. For in-hand translation, performance largely results from RAPID Hand’s natural lateral finger motion and fingertip tactile feedback, complemented by wrist-mounted vision. From a cost-effectiveness perspective, we currently refrain from deploying palm sensors, as they require complex custom designs that are expensive to implement.

[1] Yang, Shiqi, et al. "Ace: A cross-platform visual-exoskeletons system for low-cost dexterous teleoperation, 2024." Conference on Robot Learning (CoRL 2024).

[2] Cheng, Xuxin, et al. "Open-television: Teleoperation with immersive active visual feedback." Conference on Robot Learning (CoRL 2024).

[3] Lin, Toru, et al. "Learning visuotactile skills with two multifingered hands." 2025 IEEE International Conference on Robotics and Automation (ICRA 2025).

[4] Zorin, Anya, et al. "Ruka: Rethinking the design of humanoid hands with learning." Robotics: Science and Systems (RSS 2025).

[5] Yin, Zhao-Heng, et al. "DexterityGen: Foundation controller for unprecedented dexterity." Robotics: Science and Systems (RSS 2025).

[6] https://g1.haptx.com/.

[7] https://research.facebook.com/publications/linking-haptic-parameters-to-the-emotional-space-for-mediated-social-touch/

[8] https://ai.meta.com/blog/reskin-a-versatile-replaceable-low-cost-skin-for-ai-research-on-tactile-perception/

[9] https://paxini.com/

[10] https://www.xelarobotics.com/tactile-sensors

[11] Lu, Peng, et al. "Thermoformed electronic skins for conformal tactile sensor arrays." 2024 IEEE International Conference on Robotics and Automation (ICRA 2024).

[12] Yin, Jessica, et al. "Learning in-hand translation using tactile skin with shear and normal force sensing." 2025 IEEE International Conference on Robotics and Automation (ICRA 2025).