---
title: Bimanual Cable Manipulation Dataset
url: https://www.emergentmind.com/topics/bimanual-cable-manipulation-dataset
type: topic
---

# Bimanual Cable Manipulation Dataset

Bimanual cable manipulation datasets are demonstration corpora and collection stacks for learning coordinated two-arm behaviors on deformable linear or cable-like objects, including velcro cable ties, zip cable ties, and connector-insertion tasks. In the current arXiv literature, the most explicit cable-oriented instance is the ALOHA “Thread Velcro” task, which records bimanual demonstrations with multi-view RGB, two-arm joint states, and leader-arm joint actions for imitation learning on low-cost hardware. Larger bimanual resources such as YUBI and RoboCOIN broaden the surrounding ecosystem by providing large-scale dexterous or deformable-object data, but they do not define a dedicated, enumerated cable-manipulation benchmark with cable-specific counts and metrics [2304.13705] [2606.10244] [2511.17441].

## 1. Scope, public status, and corpus boundaries

The available arXiv sources describe three distinct levels of support for bimanual cable manipulation. ALOHA provides the clearest task-level coverage: its teleoperation demonstrations explicitly include threading zip cable ties and untying velcro cable ties, and its learned real-world task “Thread Velcro” is defined as inserting one end of a velcro cable tie through the loop at its other end. YUBI positions cable manipulation as part of a broader dexterous-manipulation agenda: its fingertip geometry is explicitly designed to enable “box assembly and cable insertion,” and its operational-efficiency study includes a “phone charging” task. RoboCOIN, by contrast, documents deformable-object coverage but does not explicitly report tasks involving cables, ropes, wires, or similar deformable linear objects [2304.13705] [2606.10244] [2511.17441].

| Resource | Cable-manipulation evidence | Public status relevant to cable data |
|---|---|---|
| ALOHA | “Thread Velcro,” threading zip cable ties, untying velcro cable ties | System and software are open-sourced; no explicit official downloadable cable-demonstration dataset is announced |
| YUBI | “box assembly and cable insertion,” “phone charging” | Hardware, software, and dataset stack are released; no cable-specific taxonomy or counts are reported |
| RoboCOIN | Deformable objects are covered, but no explicit cable/rope/wire tasks are named | Dataset and framework are open-sourced; no documented cable-specific subset is provided |

A recurrent misconception is that an open-source bimanual manipulation platform necessarily includes a ready-made cable dataset. The cited sources do not support that assumption. For ALOHA, the paper and website open-source the system and software to collect data, but they do not explicitly state a release of a ready-made, downloadable dataset of bimanual cable or cable-tie demonstrations. YUBI releases a large integrated stack, but the paper does not enumerate cable-specific tasks, counts, or deployment metrics. RoboCOIN publicly releases a large multi-embodiment corpus, yet the paper does not confirm a documented cable subset [2304.13705] [2606.10244] [2511.17441].

## 2. ALOHA and the explicit “Thread Velcro” dataset

Within the cited literature, ALOHA is the most direct basis for a bimanual cable manipulation dataset. Its low-cost hardware consists of two ViperX 6-DoF arms as follower robots and two smaller WidowX 6-DoF arms as leader robots. Each follower arm has a parallel-jaw gripper fitted with custom 3D-printed “see-through” fingers and gripping tape. The operator backdrives the leader arms, and the followers mirror joint positions through joint-space mapping. The setup uses no depth cameras and no external force/torque sensors; contact is handled implicitly by PID at the motor level [2304.13705].

The cable-relevant learned task in the paper is “Thread Velcro.” This task is explicitly bimanual: the left gripper holds the tie, the right gripper grasps the tail mid-air, and both arms coordinate to insert the tail through the loop. The paper reports 100 demonstrations for Thread Velcro, whereas the other real tasks—Slide Ziploc, Slot Battery, Open Cup, Prep Tape, and Put On Shoe—use 50 demonstrations each. Each demonstration episode lasts 8–14 seconds, which at 50 Hz yields 400–700 time steps per trajectory. The model is selected by lowest validation loss, but the exact train/validation split ratios are not specified; test performance is instead reported via 25 physical trials per task [2304.13705].

The ALOHA paper also describes the teleoperation context relevant to cable-like manipulation. Build time is under 2 hours with off-the-shelf parts and 3D-printed components. A protective cage of 20×20 mm aluminum extrusions with cross-cable reinforcement hosts the robot pair and cameras. Minimal calibration is required: no special encoders or precision fixtures are used, and alignment and scaling are handled in software and via manufacturer similarities between WidowX and ViperX. This suggests that the ALOHA cable-manipulation dataset is less a fixed benchmark artifact than a reproducible template for collecting task-specific bimanual demonstrations on accessible hardware [2304.13705].

## 3. Data modalities, synchronization, and bimanual action encoding

ALOHA records observations from four RGB cameras, all Logitech C922x webcams at 480×640 pixels. Two are wrist-mounted cameras, one per follower arm, providing close-up gripper views. The remaining two are a top-down camera and a front camera, the latter rotated 90° to capture vertical space. Proprioception consists of the current joint positions of both follower robots, totaling 14 DoF, with 7 per arm including gripper actuation. Actions are recorded as the leader robots’ joint positions, which are the human teleoperation signals. The paper states that choosing leader joints encodes the applied forces implicitly via the follower PID tracking error, and that this was found to be important for learning fine manipulation from demonstration [2304.13705].

Teleoperation and data recording run at 50 Hz for actions and robot state, while cameras stream at 30 fps. In practice, image observations are integrated into the 50 Hz control and data stream; because camera FPS is lower, frames are re-used or aligned by timestamps between image captures and proprioception/actions. The system records synchronized tuples of images from four views together with joint positions and actions at each control step. Precise timestamp format and file layout are not detailed in the paper, and file formats are not specified there. Typical usage is described as sequential logs of RGB frames and joint vectors at 50 Hz, with the exact format delegated to the open-source recording scripts [2304.13705].

The bimanual action space is a 14-D absolute target-joint-position vector at each time step. Rather than predicting left and right arms independently, the policy emits a single $k \times 14$ sequence chunk that jointly coordinates both arms. In the default configuration, the chunk size is $k = 100$, corresponding to approximately 2 seconds at 50 Hz. The paper states that this reduces the effective horizon by $100\times$. Coordination is therefore represented directly in the action tensor, rather than through a separate symbolic synchronization layer. This suggests that the dataset is structurally well matched to asymmetric roles such as holding, grasping mid-air, and insertion, which are central in cable-tie threading [2304.13705].

## 4. ACT: generative imitation learning over bimanual action chunks

ALOHA uses Action Chunking with Transformers (ACT), a CVAE-based imitation-learning method designed for high-precision domains in which policy errors compound over time and human demonstrations are non-stationary. Each RGB view is processed by a ResNet-18, producing $15 \times 20 \times 512$ feature maps, which are flattened to $300 \times 512$ with a 2D sinusoidal positional embedding. Four views concatenate to a $1200 \times 512$ sequence. Two additional tokens are appended: the current 14-D joint positions, projected to 512 dimensions, and the style variable $z$, also projected to 512 dimensions. A transformer encoder synthesizes the multi-view features, joints, and $z$, while a transformer decoder generates the $k$-step action sequence via cross-attention and projects it to $k \times 14$ joint targets [2304.13705].

The generative model is defined through an encoder $q_\phi(z \mid a_{t:t+k}, \bar{o}_t)$ and a decoder $\pi_\theta(\hat{a}_{t:t+k} \mid o_t, z)$. The encoder outputs a diagonal Gaussian over $z$ given the current proprioceptive observation $\bar{o}_t$ and the target action chunk $a_{t:t+k}$; the decoder predicts actions conditioned on current multi-view images, joints, and $z$. The objective combines reconstruction with a $\beta$-VAE regularizer:
$$
L = L_{\text{reconst}} + \beta \cdot D_{KL}\big(q_\phi(z \mid a_{t:t+k}, \bar{o}_t)\, \|\, N(0, I)\big)
$$
with
$$
L_{\text{reconst}} = \|\hat{a}_{t:t+k} - a_{t:t+k}\|_1.
$$
The paper specifies L1 reconstruction for precision and sets $z=0$, the prior mean, at test time for deterministic decoding [2304.13705].

ACT also uses temporal ensembling. The policy is queried at every step, and overlapping action chunks that predict the same time index are averaged with exponential weights
$$
w_i = \exp(-m \cdot i),
$$
so that for the FIFO buffer $B[t]$,
$$
a_t = \frac{\sum_i w_i \cdot B[t][i]}{\sum_i w_i}.
$$
This is intended to smooth motion while incorporating new observations continuously. The typical hyperparameters reported are hidden dimension 512, 4 encoder layers, 7 decoder layers, 8 attention heads, $\beta \approx 10$, dropout 0.1, and $k \approx 100$. The model has approximately 80M parameters, is trained from scratch per task, takes approximately 5 hours on a single 11 GB RTX 2080 Ti GPU, and runs inference at approximately 10 ms per step [2304.13705].

## 5. Performance, evaluation protocol, and practical collection guidance

For Thread Velcro, the paper reports stage-wise ACT success rates over 25 physical trials: Lift 92%, Grasp 40%, and Insert 20% final success. The baseline BeT achieves Lift 24%, Grasp 0%, and Insert 0%. The paper therefore presents ACT as substantially better than the best baseline on this task, while also making clear that the task remains difficult in the mid-air grasp and insertion phases. A related deformable task, Prep Tape, uses 50 demonstrations and yields ACT stage-wise success rates of Grasp 96%, Cut 92%, Handover 72%, and Hang 64% final success. In broader context, non-cable fine-manipulation tasks such as Slot Battery and Slide Ziploc exhibit substantially higher final success, which indicates that thin deformable manipulation is a distinct failure regime within the same system [2304.13705].

The reported failure modes for Thread Velcro are specific and informative. One failure is that the right gripper closes too early and fails to grasp the tail mid-air. Another is that insertion misses the small loop because of perception errors associated with black velcro on a black table, together with the object’s thinness, low contrast, and small projected area. The task geometry itself is narrow-tolerance: the loop is approximately 3 mm × 25 mm, and the velcro tail is approximately 2–10 mm wide depending on position along the tie. In the related user-study context of zip cable tie threading, the hole is approximately 4 mm × 1.5 mm and the tie is approximately 0.8 mm × 3.5 mm with a pointy tip [2304.13705].

The paper gives practical guidance for building cable-manipulation datasets with ALOHA. It recommends keeping teleoperation at 50 Hz; a user study shows 62% slower completion when reduced to 5 Hz ($p < 0.001$), which strongly affects fine manipulation quality and demonstration consistency. It recommends recording from all four cameras, with wrist cameras described as especially important for precise grasp and insertion. It also recommends varying initial object placement along a 15 cm line to capture corrective behavior and visual-feedback strategies, and accepting non-stationary and stochastic human strategies rather than attempting to script mid-air handovers. For environment design, it recommends a high-contrast background and good lighting to mitigate the low contrast of black velcro or cable ties on a black tabletop. The paper further states that success improves drastically as chunk size increases from $k=1$ to approximately 100, then slightly tapers if $k$ becomes too large, near open-loop [2304.13705].

## 6. Larger bimanual datasets, repurposing paths, and open problems

YUBI extends the discussion from task-specific cable demonstrations to large-scale bimanual dexterous data collection. It introduces a finger-aligned gripper with yielding, finger-driven actuation and reports a dataset of 8,434 hours across 1.20M episodes and 119 tasks. Cable-like interactions are present but not isolated as a dedicated taxonomy. The paper explicitly states that the gripper fingertip geometry is designed to enable “box assembly and cable insertion,” and its operational-efficiency experiment includes a “phone charging” task, on which YUBI achieves a 4.19× speed-up over UMI. At the same time, the paper does not report exact task names involving cable, wire, rope, cord, or hose, does not provide episode counts or hours per cable-related task, and does not report cable-specific deployment success rates. Its limitations section states that “tight cable insertion and fragile material handling” remain challenging and will need dedicated data curation and multimodal sensing [2606.10244].

YUBI’s data model is directly relevant to cable manipulation even when cable tasks are only implicit. The released dataset is converted to LeRobot format at 30 Hz and includes per-hand wrist-camera RGB, top-view stereo RGB + depth, VR-controller 6-DoF pose, gripper aperture, and task and sub-action metadata with foot-pedal segment boundaries. The paper proposes filtering task and sub-action text labels with keywords such as “cable,” “wire,” “cord,” “rope,” “hose,” “plug,” “charge,” “USB,” “connector,” “jack,” “socket,” and “outlet.” It also suggests augmenting the base dataset with explicit cable-routing or cable-insertion demonstrations collected under the same hardware and software stack. This suggests that YUBI is best understood as a scalable substrate for cable-focused subset construction rather than as a cable benchmark in its own right [2606.10244].

RoboCOIN contributes a different kind of infrastructure. It reports over 180,000 demonstrations from 15 robotic platforms, spanning 421 tasks and 16 scenarios, with a taxonomy organized by bimanual coordination patterns and object properties, plus a hierarchical capability pyramid with trajectory-level, segment-level, and frame-level annotations. Its object taxonomy explicitly includes deformable categories, and towels are given as confirmed examples. However, the paper and appendix do not list any tasks involving cables, ropes, wires, or similar deformable linear objects. The source therefore supports viewing RoboCOIN as deformable-object-capable but not cable-specific. The recommended path for cable research is to filter by deformable-object categories, mine trajectory- and segment-level language for linear-object terms, manually validate wrist-camera footage, and, if needed, add DLO-specific segment labels and offline cable-shape annotations [2511.17441].

Across these resources, the central bottleneck remains perception and contact-rich precision on thin, low-contrast deformables. ALOHA identifies visual localization error and premature gripper closure as dominant failure modes in Thread Velcro. YUBI explicitly flags tight cable insertion as challenging and points to dedicated data curation and multimodal sensing. RoboCOIN does not mention force–torque or tactile sensing and does not report cable geometry annotations such as centerlines or keypoints. The combined record suggests that current bimanual cable manipulation datasets are strongest as open, reproducible collection-and-learning stacks and weakest as standardized cable-specific benchmarks with rich annotations for flexible geometry, contact, and sub-millimeter insertion [2304.13705] [2606.10244] [2511.17441].

Source: https://www.emergentmind.com/topics/bimanual-cable-manipulation-dataset