MILE: A Mechanically Isomorphic Hand Exoskeleton and Visuotactile Robotic Hand for Data Collection in Dexterous Manipulation

Guoying Gu; Jieji Ren; Jinda Du; Ningbin Zhang; Qiaojun Yu; Xiangyang Zhu; Xingyu Wei; Yu Deng; Yufei Liu

REVIEW

MILE: A Mechanically Isomorphic Hand Exoskeleton and Visuotactile Robotic Hand for Data Collection in Dexterous Manipulation

Jinda Du, Jieji Ren, Qiaojun Yu, Ningbin Zhang, Yu Deng, Xingyu Wei, Yufei Liu, Guoying Gu

show 1 more author

Xiangyang Zhu

Robotics · Submitted 29 Nov 2025

Pith's verdict: Useful integrated hardware for dexterous data collection, but the paper oversells its validation and misses the key coupling check between human and exoskeleton.

T0 · AI review

How this review was made: An AI referee (deepseek-v4-flash) read the paper on August 3, 2026 and answered a fixed list of questions about it. T0 is the starting level every paper gets. The levels describe how deeply Pith has checked the paper, never the authors.

T0 AI review An AI read the paper and answered a fixed list of questions.
T1 Audit Automatic checks that the references and files the paper relies on exist and match.
T2 Compute The paper's code or calculations re-run with the same results.
T3 Formal statement The claims written out in a formal math language a computer can read.
T4 Proof Those claims proved, and the proof checked by computer.

How Pith checks papers →

· 5 major objections · 5 minor · Ask about this paper · The compiled record · Read the full review →

Challenge this review Comment (0) Read on arXiv ↗

Cite ▾

arXiv ID2512.00324 (version 5)

Pith ID4W5SVMNF, Pith's permanent record of this paper and its review

CitationGuoying Gu, Jieji Ren, Jinda Du, Ningbin Zhang, Qiaojun Yu, Xiangyang Zhu, Xingyu Wei, Yu Deng, Yufei Liu. "MILE: A Mechanically Isomorphic Hand Exoskeleton and Visuotactile Robotic Hand for Data Collection in Dexterous Manipulation." arXiv:2512.00324 (2025). https://pith.science/paper/4W5SVMNF

Cite this reviewBibTeX for Pith's review of this paper

Claim paper

Why Pith says this

Highlight any text to comment on it.

Editor's note

read the letter

First, the good: MILE is a genuinely useful integration. The isomorphic exoskeleton-to-hand mapping is a sensible way to avoid retargeting, the Tac-Tip sensors are compact, and the policy experiments give a credible demonstration that adding fingertip tactile input improves success on contact-rich tasks. The hardware design looks careful, and the sub-degree encoder accuracy is plausible given the MoCap comparison on the exoskeleton links. The teleoperation comparison against Manus and RealSense is a reasonable sanity check, and the user study suggests real operators can wear it.

The soft spots are real but not fatal. The important one is that the MoCap validation only tells you the encoders are accurate relative to the exoskeleton mechanism. It doesn't tell you whether the exoskeleton faithfully tracks the operator's finger joints. The paper states anthropometric scaling as a design criterion but gives no quantitative check of fit, slip, or axis misalignment across users. That is a load-bearing assumption for the 'faithful human demonstration' claim. If the exoskeleton shifts on the hand, the recorded actions are still consistent with the robot hand, but they are not the operator's intended joint angles. A simple validation with markers on the hand itself, or an IMU-based check, would go a long way.

There are also a few careless inconsistencies that should be cleaned up. The abstract promises a four-task teleoperation benchmark but Section IV.B reports three tasks; the conclusion repeats 'four in-hand tasks.' The teleoperation MaxAE is given as 5.38° in the text but 1.96° in Table I. And the 77% error reduction relative to 'potentiometer-based exoskeletons' in the conclusion is not backed by any experiment in the paper. These are fixable, but they make the paper look sloppier than it is.

The lack of released artifacts is a limitation for reproducibility. The project page exists, but no CAD or dataset is linked. For a data-collection system, the dataset is the main output; releasing even a subset would materially strengthen the contribution.

Overall, the core contribution — a mechanically isomorphic exoskeleton hand paired with tactile sensing — holds up as a useful system. It deserves a serious referee, and with the coupling validation added and the inconsistencies resolved, it would be a solid addition to the dexterous manipulation literature. I'd send it to review with a request for minor-to-major revision.

Referee report

Major objections (5)

[§IV.A.4 and Table I] Table I reports teleoperation MaxAE as 1.96°, but §IV.A.4 states that "during teleoperation ... maximum joint position error of 5.38°". These numbers are directly contradictory. Please clarify which is the correct maximum error, under what conditions it occurs, and how Table I was computed.
[Abstract, §IV.B, §VI] The abstract promises a "four-task teleoperation benchmark," and §VI claims a "64% mean gain in teleoperation success across four in-hand tasks." However, §IV.B evaluates only three teleoperation tasks: bottle-cap unscrewing, toy rotation, and ball rotation (Fig. 9a). The 64% improvement cannot be traced to the reported data. Please reconcile the task count and show the exact computation of the 64% figure.
[§V.B.2 and Table II] The abstract and conclusion state that incorporating fingertip tactile observations increases success rate by "an average of 25% over the vision-only baseline." The numbers in Table II do not support this. For ACT, the absolute success improvements across the five tasks are 4/30, 7/30, 4/30, 17/30, 2/30; for DP, they are 2/30, 3/30, 1/30, 2/30, 1/30. The average absolute gain over the ten comparisons is roughly 14 percentage points, not 25%. Relative gains vary wildly (e.g., 283% for ACT Egg Pinch). Please specify the averaging formula used and ensure it is consistent with the table.
[§III.B and §IV.A] The precision evaluation validates encoder accuracy by comparing MoCap markers attached to exoskeleton links (Fig. 4c) and the teleoperation test compares exoskeleton to robot hand via markers on both devices. Neither test verifies that the exoskeleton joint angles correspond to the operator's anatomical finger-joint angles. Given the central claim of "faithfully transfers human intent" and the use of exoskeleton commands q_h as the action signal for policy learning, the lack of any human–exoskeleton coupling validation (slip, fit, axis alignment across subjects) is a serious gap. Please add a quantitative validation that the exoskeleton readings reflect actual human joint motion, or explicitly limit the claims to exoskeleton–robot correspondence.
[§VI] The conclusion states "MILE achieves a 77% reduction in per-joint angular error relative to potentiometer-based exoskeletons," but no potentiometer-based exoskeleton baseline is described or compared anywhere in the paper. This unsupported quantitative claim should be removed or substantiated with an explicit comparison.

Minor objections (5)

[§V.A.2] The text says "we also performed 3 representative tasks" but then lists four tasks: In-Hand Rotation, Dexterous Grasp, Egg Pinch, and Cap Unscrewing. Please correct the number.
[§III.C] Typo: "The origin version of LEAPHand" should be "The original version of LEAPHand."
[§V.B.1] Typo: "instead of the the MILE-Tac hand's" should read "instead of the MILE-Tac hand's."
[Fig. 9(a)] The numeric labels in the bar chart are difficult to read in the current rendering; please enlarge or add a table with exact success rates and completion times.
[§V.B.1] The statement that applied force "is implicitly defined by the difference between" q_r and q_h is vague; clarify how this difference is used and whether it is part of the observation or action space.

How Pith checks papers → · Full audit of this paper →

The paper

The paper claims

This paper claims that a mechanically isomorphic exoskeleton and robotic hand, combined with sub-degree joint encoders and fingertip visuotactile sensors, enables retargeting-free teleoperation and produces demonstrations that improve imita

The reading

This paper tries to remove the biggest bottleneck in teaching robotic hands to manipulate objects: getting clean demonstrations of human hand motion plus the feel of contact. It presents MILE, a wearable exoskeleton wired one-to-one to a four-fingered robotic hand, so that the operator's finger-joint angles become the robot's joint commands directly, with no inverse-kinematics retargeting. The paper claims this mechanical isomorphism, together with sub-degree joint encoders and compact fingertip visuotactile sensors, yields a 64% improvement in teleoperation success over glove- and vision-based interfaces, and that adding fingertip touch to imitation-learning policies improves success by an average of 25% over vision alone. The value, if correct, is a practical pipeline for collecting synchronized visual, tactile, and joint data at scale for contact-rich dexterous manipulation.

What carries the argument

The load-bearing object is mechanical isomorphism: a one-to-one correspondence between exoskeleton and robot-hand joints (a bijection of kinematic trees with aligned axes, scaled links, and matched joint ranges) that turns teleoperation into a linear map, removing nonlinear retargeting and inverse kinematics. The second mechanism is the Tac-Tip visuotactile fingertip sensor—a deformable gel, acrylic plate, LED strip, and camera in a compact module—that records high-resolution contact deformation and shear at each fingertip, synchronized at 30 Hz with RGB-D and joint data.

Extended reading notes

Core claim

The central claim is that a mechanically isomorphic exoskeleton-robot pair makes high-fidelity dexterous data collection a hardware problem rather than an algorithmic one. Formally, with the exoskeleton and robot hand modeled as kinematic trees, isomorphism is a bijection between their joints satisfying axis alignment, link scaling, and joint-range inclusion, so teleoperation reduces to a linear joint-space map. The reported consequence is sub-degree multi-joint sensing (0.41° mean absolute error), retargeting-free operation with a peak teleoperation error of 5.38° after latency alignment, and tactile-augmented imitative policies that beat vision-only baselines by 25% on average across five

Load-bearing premise

The system assumes the exoskeleton's encoders measure the operator's true finger-joint angles, which requires a tight, slip-free mechanical coupling and a simplified kinematic model of each finger that matches every operator's anatomy; the paper states fit as a design goal but reports no quantitative check across different hand sizes.

Editorial extensions

If this is right

If the isomorphism holds, any demonstration collected on the exoskeleton is by construction within the robot hand's workspace, eliminating out-of-workspace commands and retargeting failures.
Sub-degree encoder accuracy (0.41° multi-joint mean absolute error) makes exoskeleton-derived joint commands usable directly as policy action labels, so no offline action re-estimation is needed.
Tactile observations improve policy success by 25% on average over vision-only, with the largest gap on force-sensitive tasks such as egg pinch, suggesting touch encodes contact phases that vision misses.
The same pipeline can be run by multiple operators (12-participant user study) to scale up data collection without per-operator recalibration, because joint ranges are standardized by the exoskeleton.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The paper states wearability and anthropometric scaling as design criteria but reports no quantitative fit verification across subjects, so the central claim implicitly depends on a slip-free fit; a multi-subject fit test would show whether sub-degree accuracy is operator-independent.
If the mechanical isomorphism is as tight as reported, the exoskeleton and hand could be decoupled: modular joint modules matching a different robot hand's kinematics would transfer the retargeting-free pipeline to other hands without re-deriving a mapping.
The 25% tactile gain is measured on five tasks with 30 trials each; an obvious next test is whether the gain grows with occlusion or novel objects, where vision-only policies typically degrade.
Because the hand omits the little finger and models the finger joints as simple hinges, skills requiring those joints or full-finger coupling may not be representable; extending the topology would test the upper bound of the approach.

Original abstract

Dexterous robotic hands perform complex, contact-rich manipulation. Imitation learning provides a route to such skills, but collecting human demonstrations with accurate hand actions and rich tactile information remains a key bottleneck. We present MILE, a teleoperation-based data-collection system comprising the wearable MILE exoskeleton and the mechanically corresponding MILE-Tac robotic hand. Because human-hand anatomy and wearability place tighter constraints on the high-DoF wearable, our human-first design begins with the MILE exoskeleton, equipped with custom modular joint encoders for accurate joint-angle acquisition. We then design the MILE-Tac robotic hand to share the exoskeleton's selected kinematic topology and joint-axis arrangement while satisfying robot-side implementation constraints, and equip its fingertips with compact visuotactile sensor modules. This correspondence enables direct exoskeleton-to-robot joint-space command transfer without online task-space inverse-kinematics retargeting. During teleoperation, the system synchronously records task-specific visual observations, four fingertip visuotactile streams, robot-hand proprioception, and exoskeleton-derived action commands. In a four-task teleoperation benchmark, MILE achieved a mean success rate of 76%, compared with 28% and 8% for glove-based and vision-based baselines, respectively. For downstream imitation learning, we trained paired ACT and DP policies with and without tactile input on MILE-collected demonstrations. The tactile-input variants achieved higher success rates in all paired evaluations.

The compiled record

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. Numbers in the tables are the authors', checked by code. No sentence in this record was written by a model.

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-09-23: 1 request, 6,960 input tokens.

What the abstract itself states, as probabilities: claims first 0.05 · states limitation 0.17 · human participants 0.83 · controlled comparison 0.90 · randomized assignment 0.02 · states quantitative result 0.08 · compares with prior results 0.92 · mentions code or data availability 0.31.

The abstract, sentence by sentence

arXiv listing abstract v4. Results and the main claim are underlined; show every role.

Dexterous robotic hands are expected to perform complex, contact-rich object manipulation, but learning such skills remains challenging because high-dimensional hands require high-fidelity demonstrations. Imitation learning provides a practical route for acquiring dexterous manipulation skills from human demonstrations, yet collecting synchronized multimodal demonstrations with accurate hand actions and tactile observations remains a key bottleneck. We present MILE, a teleoperation-based data-collection system comprising the human-first MILE exoskeleton and the mechanically corresponding MILE-Tac robotic hand. The system integrates custom-designed and fabricated modular joint encoders and compact MILE fingertip visuotactile sensor modules. The exoskeleton is informed by human-hand anatomy and ergonomic constraints, while the robotic hand is co-designed to preserve the selected four-finger kinematic topology. This correspondence enables joint-space command transfer and reduces reliance on task-space IK-based retargeting. The system synchronously records task-specific visual observations, four fingertip visuotactile streams, robot-hand proprioception, and exoskeleton-derived action commands. We evaluate MILE through a four-task teleoperation benchmark against representative glove-based and vision-based interfaces, and through imitation-learning experiments that compare policies trained with and without fingertip tactile input. The project page is available at https://sites.google.com/view/mile-system.

background contribution method result interpretation main claim

What the classifier reports for each sentence. The number is the probability it gave the label.
#	sentence	role	the authors' own work	states the main conclusion	strength
s0	Dexterous robotic hands are expected to perform complex, contact-rich object manipulation, but learning such skills remains challenging because high-dimensional hands require high-fidelity demonstrations.	background 0.96	0.05	0.03	asserted 0.77
s1	Imitation learning provides a practical route for acquiring dexterous manipulation skills from human demonstrations, yet collecting synchronized multimodal demonstrations with accurate hand actions and tactile observations remains a key bottleneck.	gap 0.82	0.05	0.03	asserted 1.00
s2	We present MILE, a teleoperation-based data-collection system comprising the human-first MILE exoskeleton and the mechanically corresponding MILE-Tac robotic hand.	contribution 0.99	0.91	0.12	asserted 1.00
s3	The system integrates custom-designed and fabricated modular joint encoders and compact MILE fingertip visuotactile sensor modules.	method 0.97	0.89	0.03	asserted 0.99
s4	The exoskeleton is informed by human-hand anatomy and ergonomic constraints, while the robotic hand is co-designed to preserve the selected four-finger kinematic topology.	method 0.96	0.83	0.03	asserted 1.00
s5	This correspondence enables joint-space command transfer and reduces reliance on task-space IK-based retargeting.	method 0.80	0.55	0.06	asserted 1.00
s6	The system synchronously records task-specific visual observations, four fingertip visuotactile streams, robot-hand proprioception, and exoskeleton-derived action commands.	method 0.98	0.93	0.06	asserted 1.00
s7	We evaluate MILE through a four-task teleoperation benchmark against representative glove-based and vision-based interfaces, and through imitation-learning experiments that compare policies trained with and without fingertip tactile input.	method 0.58	0.97	0.16	asserted 0.97
s8	The project page is available at https://sites.google.com/view/mile-system.	future work 0.54	0.18	0.02	asserted 0.93

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 1.00.

What was examined, and what was not

sources read

arXiv listing abstract v4

not examined

full text
figures and tables
references and citing sentences
relations to other papers

levels run

P abstract

not run

S sections
T body sentences
C citations
R cross-paper relations

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. Every number in the measurements table is the authors', checked by code. Record as JSON

Figures

**Figure 1.** Overview of MILE data collection system. The system integrates fingertip visuotactile sensing with a mechanically isomorphic MILE exoskeleton to collect dexterous hand demonstrations. It achieves sub-degree joint accuracy, enabling complex, contact-rich in-hand manipulation. A modular, low-cost tactile sensor is compact and can be integrated into system and provides high-resolution contact measurements. Policies trained on the collected data with visuotactile inputs outperform vision-only baselines on contact-rich manipulation tasks, indicating improved robustness and inference…

**Figure 2.** Size relationship among the human hand, the MILE

**Figure 3.** Exploded views of the assembly and key components.

**Figure 4.** MoCap setup and marker layouts. (a) Camera arrange

**Figure 5.** The single encoder precision test with MI(Supplementary Video 1). B. MILE Exoskeleton Guided by previous criteria, we design a human-like tree structure with per-finger serial chains [39]. The whole hand provides 17 DoFs: 5 DoFs for the thumb and 4 DoFs each for the index, middle, and ring fingers. To ensure wearability and anthropomorphic design, the dimensions of the exoskeleton were derived from the anatomy of a typical adult hand [40], as shown in

**Figure 6.** Comparison between MILE exoskeleton joint positions measured by the encoders and MoCap. The colored semi

**Figure 7.** Comparison of joint positions for the MILE exoskeleton and the MILE-Tac hand during teleoperation.

**Figure 8.** Teleoperation demonstrations with the MILE sys

**Figure 9.** (a) Teleoperation success rate and task completion time

**Figure 10.** Dataset overview: We record synchronized

**Figure 11.** The pipeline of data collection, model training and policy deployment with MILE

**Figure 12.** (a) Ablation experiment of ACT-Tac and ACT, (b)

Figures reproduced from arXiv:2512.00324 by the authors.

References

Forward citations

Cited by 5 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

SEED-UMI: Sharing the Exoskeleton between human and robot for onE-to-one Dexterous demonstration
cs.RO 2026-09 conditional novelty 7.0 of 10

SEED-UMI shares an exoskeleton between human and robot, turning demonstration collection into paired cross-embodiment supervision and enabling policies trained on raw wrist-camera images.
DexPort: A Hand-Held Skill Interface for Dexterous Hand Data Collection
cs.RO 2026-10 conditional novelty 6.0 of 10

A portable, same-hand data collection interface with a small set of finger skills improved dexterous robot policy success rates across four real-world tasks.
Koala Gripper: Co-designing Robotic Grippers and Data-Capture Devices for Scaling Dexterous Manipulation Learning
cs.RO 2026-08 conditional novelty 6.0 of 10

The Koala Gripper is a co-designed handheld capture device and robotic gripper pair with underactuated fingers and a pivoting dual thumb, shown grasping a wide range of objects and training diffusion policies for mani...
RealDexUMI: A Wearable Universal Manipulation Interface for Dexterous Robot Learning
cs.RO 2026-06 unverdicted novelty 6.0 of 10

A wearable interface with a shared dexterous hand module enables retargeting-free teleoperation and matched data collection, yielding policies with 88.75% average success across eight real-robot tasks that generalize ...
Dexterous Robotic Hands in the Foundation-Model Era: A Full-Stack Survey of Hardware, Data, Cross-Embodiment Transfer, Learning, and Evaluation
cs.RO 2026-11 conditional novelty 4.0 of 10

A full-stack survey of dexterous robotic hands concludes that human data pays off only when kinematics, cameras and contact align with the target hand, and that task-completion scores do not measure physical correctness.

How the literature uses it

2 arXiv papers cite this in 2 sentences

Each citing sentence is taken from the citing paper's LaTeX source and typed by Pith's classifier. A label appears at probability 0.8 or more (0.95 for support), a contest at 0.7 or more; the quoted words are the citing authors'. How each label was checked · JSON

Discussion

Disagree with the review, or know something it missed? Leave a comment, highlight any passage above to comment on it, or challenge this review formally.