ALOHA
Home > Hardware > Arms > ALOHA
Overview
| Item | Details |
|---|---|
| Full Name | A Low-cost Open-source Hardware System for Bimanual Teleoperation |
| Developer | Stanford IRIS Lab (Chelsea Finn) + Trossen Robotics |
| Key Researchers | Tony Z. Zhao, Zipeng Fu, Chelsea Finn |
| Publication | RSS 2023 |
| Configuration | ViperX 300 arms x 2 (Follower) + WidowX 250 arms x 2 (Leader) |
| Total Cost | Approximately $32,000 (including webcams, laptop) |
| Size | L121.9 x W101.6 x H203.2 cm |
| Applications | VLA data collection, bimanual manipulation research, imitation learning |
Key Significance
1. Standard Platform for VLA Research
ALOHA is not just hardware, but has established itself as core infrastructure for modern robot imitation learning research. At a price of $32,000, it provides performance comparable to commercial bimanual robots ($200,000+), revolutionizing research accessibility.
2. Original Platform for ACT (Action Chunking with Transformers)
ALOHA was designed as the platform for ACT algorithm development. To address the compounding error problem of simple behavior cloning, it enabled an innovative approach of predicting action sequences (action chunks) rather than single actions.
3. Major Data Source for Open X-Embodiment
In the Open X-Embodiment dataset (22 robots, 1M+ trajectories) led by Google DeepMind, ALOHA is one of the platforms providing the richest bimanual manipulation data. It serves as a key data source for RT-X model training.
4. Open Source Ecosystem
Hardware designs, software, and data collection code are all publicly available, enabling replication and extension by research labs worldwide.
Importance of Bimanual Manipulation
Human bimanual manipulation demonstrates capabilities beyond the simple sum of two arms. This originates from spatio-temporal coordination.
Tasks Requiring Bimanual Operation
| Task Type | Examples |
|---|---|
| Stabilization-Manipulation | Holding object with one arm, screwing with the other |
| Cooperative Transport | Carrying large boxes, trays |
| Tool Use | Sweeping with broom, pushing with mop |
| Cooking | Mixing ingredients, opening lids, stir-frying with spatula |
| Assembly | Tying zip ties, chain assembly |
Coordination Paradigms
- Leader-Follower: Primary arm leads the task, secondary arm supports
- Synergistic: Both arms cooperate equally for simultaneous tasks
Hardware Configuration
Complete Cost Breakdown
| Component | Cost | Notes |
|---|---|---|
| ViperX 300 x 2 + WidowX 250 x 2 | ~$9,680 | 2 followers + 2 leaders set |
| Cameras, sensors, mounts | ~$5,000 | Wrist cameras + external views |
| Laptop (with GPU) | ~$5,000 | Consumer-grade GPU |
| Frame and other hardware | ~$12,320 | Aluminum extrusion, cables, power, etc. |
| Total | ~$32,000 | Can reduce further with 3D printing |
Note: Costs based on original paper and Trossen Robotics pricing (2023). May vary by exchange rate and purchase timing.
ViperX 300 Specifications
| Spec | Value |
|---|---|
| Degrees of Freedom | 6 DoF |
| Reach | 750mm |
| Payload | 750g |
| Motors | DYNAMIXEL X-Series |
| Waist/Shoulder | XM540-W270 (4096 level feedback, ±0.1mm repeatability) |
| Wrist/Gripper | XM430-W350 (enhanced thermal management) |
Camera System
| Version | Camera | Features |
|---|---|---|
| ALOHA (Original) | Consumer webcams | RGB, multiple viewpoints |
| ALOHA 2 | Intel RealSense D405 | Wide FOV, depth, global shutter |
Gripper
- Follower: Parallel gripper with enhanced grip tape
- Leader (ALOHA 2): Replaced with XC430-W150-T motor (low-friction metal gears)
Teleoperation Method
ALOHA’s teleoperation uses backdriving-based puppeteering.
Operating Principle
User → Physical manipulation of leader arm → Read joint positions → Synchronize follower arm- Physical Backdriving: User directly moves leader arm (WidowX) by hand
- Real-time Synchronization: Leader joint positions immediately reflected to follower (ViperX)
- Data Collection: RGB images + joint states recorded simultaneously
Key Advantages
| Advantage | Description |
|---|---|
| Intuitiveness | Natural interface of directly moving robot arm |
| Low Latency | Lower latency compared to joystick/VR controllers |
| Force Feedback | Sensing physical resistance enables delicate manipulation |
| Low Cost | Uses only existing arm encoders without additional sensors |
Required Environment
- 6+ USB3 ports (4 robots + 2 cameras)
- Connection instability possible when using USB hub
ALOHA Version Comparison
| Item | ALOHA (Original) | ALOHA 2 |
|---|---|---|
| Release | 2023 (RSS) | 2024 |
| Gripper | High friction | Low-friction rail design |
| Gravity Compensation | Rubber bands | Passive kinematic mechanism |
| Camera | Webcam | Intel RealSense D405 |
| Frame | Basic | 20x20mm aluminum extrusion |
| Durability | Moderate | Enhanced |
Extension to Mobile ALOHA
Mobile ALOHA is an extended version with ALOHA mounted on a mobile base (AgileX Tracer).
| Item | Details |
|---|---|
| Base | AgileX Tracer AGV (~$7,000) |
| Movement Method | User physically connected to system, backdrivin wheels |
| Data | Base velocity + arm puppeteering recorded simultaneously |
| Learning | Autonomous execution possible with 50 demonstrations |
Mobile ALOHA Training Task Examples
- Putting pot in cabinet
- Calling elevator
- Pushing in chair
- Stir-frying shrimp
- Cleaning wine spill
- High-five
See Mobile ALOHA for details.
VLA Research Applications
ALOHA is a key evaluation/training platform for various VLA (Vision-Language-Action) models.
ACT (Action Chunking with Transformers)
| Item | Details |
|---|---|
| Core Idea | Predict action sequences (chunks) instead of single actions |
| Architecture | Conditional VAE + Transformer Encoder/Decoder |
| Problem Solved | Reduces compounding error by factor of k (k = chunk length) |
| Performance | 80-90% success rate with 10-minute demonstration |
| Tasks | Fine-grained bimanual manipulation like opening transparent cups, battery insertion |
OpenVLA
| Item | Details |
|---|---|
| Parameters | 7B |
| Base | Llama 2 + DINOv2 + SigLIP |
| Training Data | 970k real robot demonstrations (including ALOHA) |
| Performance | 16.5% higher success rate than RT-2-X (55B) |
Physical Intelligence Pi Series
| Model | Features |
|---|---|
| Pi-0 | 3B PaLiGemma VLM + 300M Diffusion action expert |
| Pi-0-FAST | Speed improved with tokenized action output |
| Pi-0.5 | Open-world generalization through heterogeneous data co-training |
Pi-0 is benchmarked on various robot platforms including ALOHA, significantly outperforming existing baselines like OpenVLA, Octo.
OpenVLA-OFT
Achieves high-frequency language-based control with 7B VLA policy on ALOHA. 97.1% success rate on LIBERO benchmark, surpassing Pi-0, Diffusion Policy, etc.
Software Ecosystem
| Package | Details |
|---|---|
| ROS / ROS 2 | Drivers, URDF, Gazebo simulation |
| MoveIt | Motion planning support |
| LeRobot | Hugging Face robot learning library integration |
| ACT Code | Official training/inference code released |
References
Papers
- Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware (ACT, RSS 2023)
- ALOHA 2: An Enhanced Low-Cost Hardware for Bimanual Teleoperation
- Mobile ALOHA: Learning Bimanual Mobile Manipulation with Low-Cost Whole-Body Teleoperation
- Open X-Embodiment: Robotic Learning Datasets and RT-X Models
- OpenVLA: An Open-Source Vision-Language-Action Model
- Pi-0: A Vision-Language-Action Flow Model for General Robot Control
Project Sites
- ALOHA Official Page
- ALOHA 2 Official Page
- Mobile ALOHA Official Page
- Stanford IRIS Lab
- Open X-Embodiment