N/A
ALPHA
ALPHA, developed by TeknTrash, is a humanoid robot designed to transform waste handling by automating repetitive, unsafe tasks in recycling facilities. With advanced AI, 6‑DOF arms, 3D vision, and cloud connectivity, it can sort, lift, and navigate autonomously to increase efficiency and reduce contamination in waste streams.
Description
ALPHA (Automated Litter Processing Humanoid Assistant), developed by UK-based TeknTrash Robotics, is a specialized humanoid robot engineered to revolutionize waste handling in recycling facilities by automating repetitive, hazardous, and unsanitary tasks traditionally performed by humans. Standing at 200 cm tall and weighing 90 kg, ALPHA features a modular architecture comprising a wheeled mobile base for autonomous navigation, a lifting column for height adjustment (up to 2000 mm, 150 mm/s speed), dual 6-DOF arms with grippers capable of lifting 5 kg total payload, a head unit with 2-DOF binocular depth camera for global object identification and obstacle avoidance, and a chest-mounted Nvidia Jetson Orin AGX 64 compute module delivering 275 TOPS of AI processing power via 2048 CUDA cores, 64 Tensor Cores, 12-core Arm Cortex-A78AE CPU, and integrated accelerators (2x NVDLA, PVA v2.0). The robot's sensory suite includes hyperspectral cameras for material identification beyond visible light (UV, IR, X-ray spectra), arm-mounted 3D cameras (85° H x 58° V FOV, 0.25-2.5 m depth), base lidar, depth camera, ultrasonic sensors (1 safety edge + 2 units), IMU/encoder for navigation, and a monocular camera/microphone on the lifting column. Operating on Ubuntu 24 with WiFi cloud connectivity, ALPHA supports team-based operations where a central conveyor belt camera allocates tasks via Google Cloud infrastructure, offloading heavy computation (e.g., 6+ MB/s hyperspectral data per robot) to Compute Engine GPUs for real-time inference, reducing on-board latency to ~20 ms glass-to-action.
AI architecture leverages Nvidia Isaac GR00T Vision-Language-Action (VLA) models trained on vast datasets from HoloLab, TeknTrash's VR-based data capture system using Meta Quest 3 headsets to record operatives' motions in LeRobot format. Over 49 TB of structured/unstructured data (hyperspectral images, motion logs, video streams) is processed via Cloud Data Fusion for model fine-tuning, enabling ALPHA to mimic human picking (30-40 items/min) without fatigue, achieving up to 95% material purity vs. 75% human rates and reducing contamination by 25%. Cloud-native design supports Robots-as-a-Service (RaaS), remote firmware updates, and scalability to 1,000+ sites.
Real-world deployment began in 2025 with a pilot at Sharp Group's Rainham, East London facility (2,800 tonnes/week throughput), addressing labor shortages for 24/7 ops. HoloLab captured real-time sorting data from workers, training ALPHA for hyperspectral waste tracking, dexterous picking (grippers vs. suction), and autonomous mobility. Outcomes include 10%+ recovery efficiency gains, granular EPR compliance data, and injury reduction in a sector with 17x higher fatalities. As a prototype (IP32 ingress, 7-hour runtime), ALPHA outperforms fixed robotic arms by enabling free movement to retrieve missed items, positioning it for full facility takeover in waste processing, carrying, and policing.
Key Features
Hyperspectral Vision
Advanced cameras capture UV, IR, and X-ray spectra for accurate identification of dirty or incomplete waste items, enabling early tracking from conveyor starts.
Dual 6-DOF Dexterous Arms
Equipped with grippers (4 fingers, 2 DOF hands) for reliable picking without constant cleaning, supporting 5 kg payload and human-like manipulation.
Cloud-Coordinated Team Operation
Permanent WiFi connectivity allows central task allocation to robot teams, offloading compute to Google Cloud for scalable, low-latency performance.
Autonomous Wheeled Navigation
Lidar, depth cameras, ultrasonics, and IMU enable map building, obstacle avoidance, positioning, and auto-charging on dynamic factory floors.
Nvidia Isaac GR00T AI Training
VLA models trained via HoloLab VR data capture human motions precisely, supporting 24/7 fatigue-free sorting at 30-40 picks/min.
Specifications
| Availability | Prototype |
|---|---|
| Nationality | UK |
| Website | https://www.tekntrash.com/ |
| Degrees Of Freedom, Overall | 13 |
| Degrees Of Freedom, Hands | 2 |
| Height [Cm] | 200 |
| Manipulation Performance | 2 |
| Navigation Performance | 2 |
| Max Speed (Km/H) | 0.9 |
| Strength [Kg] | 5 |
| Weight [Kg] | 90 |
| Runtime Pr Charge (Hours) | 7 |
| Safe With Humans | Yes |
| Cpu/Gpu | Nvidia Orin AGX 64 |
| Ingress Protection | IP32 |
| Camera Resolution | 640×480 |
| Connectivity | WiFi (cloud via Google Compute Engine) |
| Operating System | Ubuntu 24 |
| Llm Integration | deepseek |
| Latency Glass To Action | 20 ms |
| Motor Tech | Realman |
| Gear Tech | Realman |
| Main Structural Material | Aluminum, Plastic |
| Number Of Fingers | 4 |
| Main Market | Recycling |
| Verified | Not verified |
| Height | 200 cm |
| Weight | 90 kg |
| Degrees Of Freedom (Overall) | 13 |
| Arms | Dual 6-DOF with 2-DOF grippers (4 fingers) |
| Payload | 5 kg total |
| Max Speed | 0.9 km/h |
| Runtime | 7 hours per charge (3-hour recharge) |
| Processor | Nvidia Jetson Orin AGX 64 (12-core Arm Cortex-A78AE, 2048 CUDA cores, 64 Tensor cores, 275 TOPS, 64GB eMMC, 1TB SSD) |
| Os | Ubuntu 24 |
| Sensors | Vision: Hyperspectral cameras (UV/IR/X-ray), Arm 3D cameras (640x480, 85°x58° FOV, 0.25-2.5m depth), Head binocular depth, Base depth/LiDAR, Lifting monocular + mic, Proximity: Ultrasonic (1 safety edge + 2 units), IMU/encoder |
| Navigation | Deep learning camera fusion, SLAM, dynamic obstacle avoidance |
| Ingress | IP32 |
| Materials | Aluminum frame, plastic components |
| Motors/Gears | Realman |
| Ai Models | Nvidia Isaac GR00T VLA, deepseek LLM |
| Data Rate | 6+ MB/s per robot (hyperspectral) |
| Thermal | N/A (passive cooling implied via Orin design) |