PaperlayerView exact revision on arXiv

Source-linked AI summary

ALOHA 2: An Enhanced Low-Cost Hardware for Bimanual Teleoperation

ALOHA 2 Team, Jorge Aldaco, Travis Armstrong, Robert Baruch, Jeff Bingham, Sanky Chan, Kenneth Draper, Debidatta Dwibedi, Chelsea Finn, Pete Florence, Spencer Goodrich, Wayne Gramlich, Torr Hage, Alexander Herzog, Jonathan Hoech, Thinh Nguyen, Ian Storz, Baruch Tabanpour, Leila Takayama, Jonathan Tompson, Ayzaan Wahid, Ted Wahrburg, Sichun Xu, Sergey Yaroshenko, Kevin Zakka, Tony Z. Zhao

arXiv:2405.02292v1cs.ROcs.LG

TL;DR

Scaling data collection for complex bimanual manipulation motivates ALOHA 2's hardware and simulation redesign. The authors report easier teleoperation of challenging tasks and hundreds of demonstrations per robot per day, alongside improved performance, user-friendliness, and robustness versus the original system.

  • Problem

    Scaling ALOHA data collection for complex manipulation calls for improvements in task range, operator ergonomics, and system robustness.

    Source: ALOHA 2 · p. 2

  • Method

    ALOHA 2 combines redesigned grippers, passive gravity compensation, a simplified frame, smaller D405 cameras, and a system-identified MuJoCo model.

    Source: ALOHA 2 · p. 2, Simulation · p. 7

  • Results

    ALOHA 2 improves performance, user-friendliness, and robustness over the previous system, and makes challenging tasks easier to teleoperate with hundreds of demonstrations collected per robot per day.

    Source: ALOHA 2 · p. 2, Conclusion · p. 9

  • Takeaways & Limitations

    The authors hope ALOHA 2 will enable large-scale data collection for fine-grained bimanual manipulation to advance robot learning research.

    Source: Conclusion · p. 9

Abstract

from arXiv · show

Diverse demonstration datasets have powered significant advances in robot learning, but the dexterity and scale of such data can be limited by the hardware cost, the hardware robustness, and the ease of teleoperation. We introduce ALOHA 2, an enhanced version of ALOHA that has greater performance, ergonomics, and robustness compared to the original design. To accelerate research in large-scale bimanual manipulation, we open source all hardware designs of ALOHA 2 with a detailed tutorial, together with a MuJoCo model of ALOHA 2 with system identification. See the project website at aloha-2.github.io.

ALOHA 2

ALOHA 2 redesigns the bimanual teleoperation platform to support larger-scale, more diverse data collection, targeting task range, operator ergonomics, and robustness. Its hardware and simulation changes are reported to make challenging tasks easier to teleoperate and support hundreds of demonstrations per robot per day.

ALOHA 2

  • The workcell pairs two ViperX 6-DoF follower arms with two smaller, kinematically similar WidowX leader arms, which users backdrive to teleoperate the followers.Cameras capture RGB data from multiple viewpoints during teleoperation.Source: ALOHA 2 · p. 2
  • ALOHA 2 targets scaling data collection by improving task range, user friendliness and ergonomics, and system robustness.The scaling aims include more robots, more hours of data per robot, and greater diversity in data collection.Source: ALOHA 2 · p. 2
  • Hardware changes add low-friction gripper rails, passive gravity compensation, a simplified frame, and smaller D405 cameras; renderings show the follower camera mount and leader swappable finger mounts.The cameras provide depth and global shutter, while the simplified frame retains rigid camera mounting points.Source: ALOHA 2 · p. 2, Figure 2 · p. 3
  • The team models ALOHA 2's exact specifications in MuJoCo to support data collection, policy learning, and evaluation in simulation for challenging manipulation tasks.Source: ALOHA 2 · p. 2
  • The upgrades make challenging tasks easier to teleoperate and support collecting hundreds of demonstrations per robot per day.Examples include folding a T-shirt, tying a knot, throwing objects, and industrial tasks with tight tolerances.Source: ALOHA 2 · p. 2

Hardware

ALOHA 2 redesigns its grippers, gravity compensation, frame, and camera placement to improve teleoperation and support broader data collection. A six-user comparison favored passive over active gravity compensation on a shape-insertion task.

Leader Grippers (The device the human holds in their hands)

  • The low-friction leader rail replaces the scissor design and requires approximately 10 times less force to open and close.Lower friction reduces hand fatigue and strain during long sessions, notably in the lumbrical muscles used to open the gripper.Source: Leader Grippers (The device the human… · p. 3
  • In a six-user candy-unwrapping study, nearly all operators rated the linear rail design well, despite varying preferences across designs.The alternatives were the original scissor design and a spring-loaded trigger design.Source: Leader Grippers (The device the human… · p. 3, Figure 3 · p. 4

Follower Grippers (The end-effector of the robot, i.e. the robot fingers that manipulate objects)

  • Low-friction follower grippers reduce perceived leader-follower latency and provide twice the force of the previous design for stronger, more stable grasps.The mechanism uses deformable 3D-printed carbon-fiber nylon in place of the original PLA and acrylic structure, improving safety under load.Source: Follower Grippers (The end-effector of… · pp. 3–4

Gravity Compensation

  • The passive leader-arm gravity system uses off-the-shelf adjustable hanging retractors, allowing operators to tune load balancing to their comfort.The authors also report that it can permit safer teleoperation with leader joint actuation disabled and naturally center the arms.Source: Gravity Compensation · pp. 4–5
  • 1.38 vs. 0.97 shapes inserted per minute: operators averaged more with passive than active gravity compensation.The comparison used six users in 10-minute sessions; the authors speculate that passive compensation provides smoother, more predictable movements.Source: Gravity Compensation · pp. 4–5

Grippers

  • 14.68N to 0.84N operator opening force; follower tip closing force is 27.9N vs. 12.8N.The figure reports lower operator force for ALOHA 2 and more than double the follower closing force compared with the old design.Source: Figure 4 · p. 5

Frame

  • The simplified aluminum-extrusion frame removes the far-side vertical supports, leaving room for human-robot interaction and larger props.It supports the leader robots and gravity compensation system and provides mounts for overhead and worms-eye cameras.Source: Frame · p. 5, Figure 5 · p. 6

Cameras

  • Lower-profile wrist cameras reduce collision states and improve teleoperation for fine-grained tasks involving close contact or tight spaces.The upgraded RealSense D405 cameras provide high-resolution RGB and depth in a small form factor, with global shutter.Source: Cameras · pp. 5–6

Teleoperation

The teleoperation workflow uses ROS2, leader and follower arms, and foot pedals to manage collection sessions. It logs synchronized sensor data and applies checks intended to keep downstream datasets complete and reliable.

Session workflow

  • The ROS2 workflow starts both arms at home; operators begin a session by closing either leader gripper and save or discard it with foot pedals.The system logs images, leader and follower joint positions, and auxiliary ROS2 data, with joint data recorded at 50Hz.Source: Teleoperation · p. 6

Data logging

  • Session statistics expose sensor availability and latency during collection, and sessions shut down automatically when data is missing.Logged records include operator username, time, and robot identifier to support downstream filtering if issues arise.Source: Teleoperation · p. 6

Camera views

  • Four views are recorded during teleoperation: overhead, worms-eye, left wrist, and right wrist, each at 848 x 480 RGB.The wrist cameras complement the scene views with perspectives from both arms.Source: Figure 6 · p. 7

Ergonomics

  • Ergonomic guidance recommends breaks of at least two minutes every ten minutes and varying tasks to avoid repeated continuous motions.The guidance also lists low-friction grippers, passive gravity compensation, swappable finger attachments, and adjustable-height chairs.Source: Teleoperation · p. 7

Simulation

ALOHA 2’s MuJoCo workcell model is calibrated against real trajectories and designed for higher physical and visual fidelity. It supports scalable simulated teleoperation and data collection.

Simulation

  • The MuJoCo model is more physically accurate and visually faithful than earlier ALOHA models, with physics parameters fitted to 11 real trajectories using nonlinear least squares.The optimization minimizes residuals between real and simulated trajectories.Source: Simulation · p. 7
  • The model tunes actuator gains, damping, armature, joint friction, and torque limits, while matching camera intrinsics and importing workcell and gripper assets.The gripper is represented by a position-controlled linear actuator with an equality constraint between its fingers.Source: Simulation · p. 8
  • The model includes ViperX robots, wrist cameras, and a replica of the aluminum-extrusion frame, with robot and camera positions modeled to match ALOHA 2.Figure 7 depicts the simulated workcell.Source: Figure 7 · p. 8
  • Simulation collection runs show four camera views of teleoperated tasks using objects from the Google Scanned Objects Dataset (Downs et al., 2022).The realistic model supports fast, intuitive, scalable collection with an ALOHA 2 WidowX leader setup.Source: Figure 8 · p. 8, Simulation · p. 9
  • The authors hope the open-source, system-identified model will support sharing teleoperated simulation data across institutions and accelerate policy-learning research in simulation.Source: Simulation · p. 9

Conclusion

ALOHA 2 is presented as a low-cost bimanual teleoperation system with improvements to performance, user-friendliness, and robustness. Its hardware changes and high-quality simulation model are intended to support large-scale data collection for fine-grained bimanual manipulation.

Conclusion

  • ALOHA 2 is a low-cost bimanual teleoperation system designed to improve performance, user-friendliness, and robustness over the previous ALOHA system.Source: Conclusion · p. 9
  • Hardware improvements target grippers, gravity compensation, the frame, and cameras, alongside a high-quality simulation model.Source: Conclusion · p. 9
  • The authors hope ALOHA 2 will support large-scale data collection for fine-grained bimanual manipulation and advance robot-learning research.Source: Conclusion · p. 9

Related papers

Loading 2405.02292v1…