Inside the SVRC Mountain View Robotics Lab: Infrastructure, Fleet, and Research

The Facility

The SVRC Mountain View lab is located at 1117 Independence Avenue, Mountain View, CA 94043 — in the heart of Silicon Valley's robotics corridor, with proximity to Google, Waymo, and numerous robotics startups. The facility covers approximately 3,500 square feet, organized into four distinct zones optimized for different functions.

Zone 1 — Data Collection Floor (1,800 sq ft): The main production area houses 12 active data collection stations, each with a standardized camera rig, lighting setup, and workspace configuration. Stations are spaced 8 feet apart to prevent inter-station interference (lighting crosstalk, object migration between workspaces). The floor is epoxy-coated level concrete — essential for consistent robot base positioning and repeatability across sessions. Each station has a dedicated 20A power circuit and hardwired gigabit Ethernet to the central NAS.

Zone 2 — GPU Compute Room (400 sq ft): Climate-controlled server room housing the training cluster, NAS array, and network infrastructure. Maintained at 68F/20C with redundant cooling. The room is on a dedicated 60A three-phase circuit with UPS backup for graceful shutdown.

Zone 3 — Hardware Integration Lab (800 sq ft): Where new robot platforms are assembled, configured, calibrated, and tested before deployment to the collection floor. Includes a dedicated workbench with soldering station, 3D printer (Prusa MK4S) for custom brackets and fixtures, and a full set of metric and imperial tooling. This is where new hardware is validated before it enters production data collection.

Zone 4 — Conference and Collaboration (500 sq ft): Meeting space with 4K display for remote collaboration, capable of streaming live robot feeds from any station on the collection floor. Used for customer demos, research presentations, and remote data review sessions.

Robot Fleet

The active robot fleet includes 20+ robots across several platform categories, selected to cover the broadest range of data collection scenarios demanded by our customers.

Single-Arm Manipulation Stations (8 units)

Our standard manipulation platform is the OpenArm 101, SVRC's own open-source 6-DOF arm ($4,500 per unit). Each station is equipped with:

• OpenArm 101 follower arm with integrated gripper (2 kg payload, 850mm reach)
• OpenArm leader arm for teleoperation (passive kinematic replica)
• Wrist-mounted Intel RealSense D405 (close-up manipulation view, 640x480 at 30fps)
• Overhead Intel RealSense D435i (workspace context, 1280x720 at 30fps)
• Side-view Basler acA1920-40gc (additional perspective, hardware-triggered sync)
• ATI Mini45 6-axis force/torque sensor (wrist-mounted, 1kHz sampling)
• Dedicated mini-PC (Intel NUC with RTX 4060 mobile, Ubuntu 22.04, ROS 2 Humble)

All three cameras are synchronized to within 1ms via hardware trigger. Frame rate: 30fps standard, 60fps available for high-speed tasks. These stations handle the majority of standard manipulation task collection — pick-and-place, insertion, assembly, and tool use tasks.

The OpenArm's open-source design means we can modify mounting configurations, swap grippers, and add sensors without vendor restrictions. We maintain a library of 3D-printed gripper fingers optimized for different object categories (cylindrical, flat, deformable).

Bimanual Stations (2 units)

Full bimanual setups configured for two-handed manipulation tasks. The DK1 bimanual system provides 14-DOF total (two 7-DOF arms), with overhead camera plus wrist cameras on each arm. These stations are used for:

• Bimanual assembly (holding + inserting, opening containers)
• Garment handling (folding towels, sorting laundry)
• Dual-object manipulation (pouring from one container to another)
• Two-handed tool use (holding a workpiece + using a tool)

Bimanual data collection requires significantly more operator skill than single-arm tasks. Our bimanual operators have a minimum of 40 hours of training before they collect production data. The extra training investment pays off in higher-quality demonstrations that produce better policies.

Dexterous Hand Stations (2 units)

Single-arm plus multi-finger dexterous hand for in-hand manipulation, tool use, and high-dexterity tasks. Currently configured with Inspire Hands RH56DP (6-DOF per finger, tactile sensing at fingertips). Teleoperation via Paxini tactile glove for natural finger motion mapping.

Dexterous manipulation data collection is the most challenging and highest-value category in our catalog. Each dexterous demonstration takes 2-3x longer than a standard gripper demonstration and requires specialized operator training. The resulting data commands premium pricing but is in high demand from teams training dexterous manipulation policies. See our dexterous hands comparison for more on the hardware landscape.

Mobile Manipulation Platforms (2 units)

Hello Robot Stretch 3 for navigation + manipulation tasks, and a Mobile ALOHA configuration for bimanual mobile manipulation research. These platforms enable data collection for tasks that span multiple rooms or require the robot to navigate to objects before manipulating them — a critical capability for household and healthcare applications.

Mobile manipulation data collection requires a structured environment with defined navigation waypoints, consistent lighting across rooms, and motion capture or SLAM-based localization for accurate pose tracking. Our lab's open floor plan with configurable partitions allows us to create repeatable multi-room scenarios.

Specialty Platforms

• Unitree G1 humanoid (1 unit): Available for locomotion data collection, whole-body manipulation research, and customer demonstrations. The G1 stands 127cm tall, weighs 35kg, and has 23 DOF. It walks at up to 2 m/s on flat ground and can carry objects while walking. SVRC is one of the few facilities in the US with a G1 available for third-party research use. Leasing available.
• Kitchen manipulation station: A dedicated workspace with real kitchen fixtures (sink, countertop, cabinet with drawers) for collecting household manipulation demonstrations. The arm (Fanuc CR-7iA/L, 7kg payload) is mounted to simulate a kitchen-deployed robot. Used for tasks like loading dishes, opening drawers, and pouring liquids.
• Glove teleoperation station: A dedicated station for high-dexterity tasks using teleoperation gloves (Paxini, Manus) mapped to dexterous hands. Includes forearm EMG sensors for intent detection and a 6-camera array for high-coverage recording.
• Orca Hand integration station: For testing and data collection with the Orca Hand, a high-DOF dexterous end-effector designed for research-grade in-hand manipulation.

Data Collection Capacity and Throughput

Understanding our throughput helps customers plan data collection timelines and budgets accurately.

These numbers reflect sustained production throughput, not peak burst rates. We plan capacity conservatively because operator fatigue degrades data quality — we enforce mandatory breaks every 45 minutes and limit collection shifts to 6 hours to maintain consistent demonstration quality throughout the day.

Data Pipeline and Quality Infrastructure

Raw teleoperation recordings are not training data. Our automated pipeline transforms them into quality-certified datasets.

• Episode recording: Custom HDF5 pipeline captures synchronized observations (joint states at 50Hz, camera frames at 30fps, F/T readings at 1kHz downsampled to 50Hz) and action sequences. Compatible with LeRobot, RLDS, and custom loaders.
• Automated success classification: Fine-tuned ResNet-50 on the final 10 frames of each episode. Achieves 92% accuracy on held-out test sets for binary success classification. All new episodes pass through the classifier; borderline predictions (confidence 0.4-0.7) are queued for human review.
• Trajectory smoothness scoring: Computes jerk metric (third derivative of joint positions) and flags episodes with smoothness below the 30th percentile for the task type. Jerky demonstrations are either discarded or queued for operator feedback.
• Diversity coverage analysis: Tracks object position coverage, orientation diversity, and visual embedding space coverage across the dataset. Generates heatmaps showing which areas of the workspace and which object configurations are under-represented, enabling targeted additional collection.
• Format validation: Schema validation at write time catches format errors, missing fields, and timestamp misalignments before episodes enter the dataset. This eliminates the "crash on episode 3,847 during training" failure mode.

Dataset Library: 50TB+ and Growing

SVRC has accumulated over 50TB of robot demonstration data across 3 years of continuous collection. The dataset library is indexed in PostgreSQL by task type, robot platform, operator, date, success label, and quality scores.

Portions of this library are available to customers with appropriate data access agreements. Custom task collection is available through our data services program.

GPU Cluster

The compute infrastructure consists of 8x NVIDIA A100 80GB GPUs across two servers, with 25GbE interconnect to the NAS storage array. The cluster is shared between internal research and customer training jobs, managed via a job queue with SLA tiers:

Typical training turnaround for standard policy sizes: ACT at 200 demos completes in 4-6 hours on a single A100. Diffusion Policy at 500 demos completes in 8-12 hours. Large foundation model fine-tuning runs (100K+ demos, large backbone) are scoped individually and may use multi-GPU configurations.

Customers receive Weights & Biases training dashboards for real-time monitoring of their training runs, including loss curves, success rate on held-out evaluation episodes, and per-task performance breakdowns.

The SVRC Data Platform

The SVRC data platform is the software layer that ties together hardware, data, and training. Key capabilities:

• Episode recording and management: Record, tag, annotate, and manage episodes directly in the browser. Stream live camera feeds and joint state from any connected device.
• Dataset explorer: Browse, filter, and visualize episodes by task type, robot platform, success label, quality score, and collection date. Preview episode playback without downloading raw files.
• Quality dashboard: Per-dataset quality scores across the six dimensions (diversity, consistency, completeness, accuracy, balance, format). See our data quality framework for details.
• Export to training formats: One-click export to LeRobot HDF5, RLDS, or custom formats. Exports include all metadata required for training (camera calibration, robot URDF, task descriptions).
• Training job submission: Submit training jobs to the GPU cluster from the platform interface, with pre-configured recipes for ACT, Diffusion Policy, and OpenVLA fine-tuning.

Lab Access Programs

SVRC offers several ways to access the Mountain View lab and its resources, depending on your needs and stage.

Facility Visit

Schedule a visit to tour the lab, see the robot fleet in action, and discuss your project with our engineering team. Visits are free and typically last 1-2 hours. Contact us to schedule.

Desk and Station Rental

For teams that want to work on-site with their own hardware or with SVRC hardware, we offer desk and station rental by the day, week, or month. Station rentals include access to a configured collection station with cameras, compute, and the SVRC platform. Contact us for current rates.

Managed Data Collection

The most common engagement model. You define the task, SVRC designs the collection protocol, assigns trained operators, runs the collection, quality-filters the data, and delivers a certified dataset. Pricing starts at $2,500 for a pilot (50 demos) and $8,000 for a standard campaign (500 demos). See data services for the full menu.

Hardware Leasing

Lease SVRC hardware for use at your own facility. Available platforms include OpenArm 101 ($800/month), DK1 bimanual ($1,500/month), and select commercial arms. Leasing includes remote setup support and access to the SVRC platform. See leasing for details.

East Coast: Allston, MA

SVRC's East Coast facility is located at 125 Western Avenue, Allston, MA 02134 — adjacent to the Harvard School of Engineering and Applied Sciences. The Allston lab provides a smaller-scale complement to the Mountain View facility, with 4 active collection stations and access to the Boston-area robotics research community (MIT CSAIL, Harvard SEAS, Northeastern). East Coast customers benefit from reduced shipping time and the option of on-site visits without cross-country travel.

Collaboration Models

We are actively looking for research partners in dexterous manipulation, mobile manipulation, and foundation model fine-tuning. If you are interested in visiting the lab or starting a collaboration, see our join page or reach out via the contact form.

How to Get Started

1. Schedule a lab visit — see the facility, meet the team, discuss your project. Contact us.
2. Start with a pilot — $2,500 for 50 quality-certified demonstrations of your task. Evaluate the data quality before committing to a larger campaign.
3. Scale to production — $8,000+ for full campaigns with guaranteed quality, delivery timelines, and format compatibility with your training pipeline.