Robotics & Embodied AI
VXI-IR5 is an on-device large-model co-processor built for robots. It works alongside your existing robot controller and adds local language understanding, visual semantics and task reasoning within a tight power and thermal budget. Our own engineers stay with you from model import to a robot in the field.
Join the Early Access Program · Estimate my model's needs
Source: International Federation of Robotics (IFR), World Robotics 2026 press release, September 24, 2026. Read the IFR release
Putting large models into robots is limited by the machine itself, not by headline compute.
Battery life, enclosure cooling and fan noise decide how much compute fits on board. Efficiency matters more than peak numbers.
Round-trip latency and unstable links disrupt navigation, obstacle handling and human interaction. Key decisions belong on the robot.
Factory video, warehouse layouts and customer sites often must not leave the premises. Local inference keeps data governance simple.
VXI-IR5 joins your existing architecture as a co-processor: the controller keeps motion control and safety, VXI takes the large-model inference.
The two exchange inference results over a host interface, so your control path and safety design stay as they are.
The exact interface, operating system and integration approach are confirmed with you during evaluation, based on your controller platform.
Every token a large model generates needs a large share of its weights. When weights sit in external memory, bandwidth sets the ceiling.
Take a 1.5-billion-parameter model quantized to 4 bits: about 0.75 GB of weights. If every token reads all weights once from external memory at 25.6 GB/s, the theoretical ceiling is about 34 tokens per second — before any other overhead.
VXI uses SRAM compute-in-memory to keep frequently accessed data next to the compute units, cutting external traffic per token. The efficiency comes from data-path design, not from the most advanced process node.
Each has clear deployment demand today and a strong need for local decisions within a limited power budget.
On-site need：Recognize goods, people and obstacles in changing environments and understand dispatch instructions.
What VXI adds：Local visual semantics and natural-language instructions, with less dependence on cloud dispatch.
On-site need：Long patrols in substations, utility tunnels and plants, often with unreliable networks.
What VXI adds：Offline anomaly recognition and patrol summaries — send back results instead of raw video.
On-site need：Move between stations and pick and place, understanding parts, workstations and work instructions.
What VXI adds：Visual semantics and task planning support decisions, while control and safety stay with the controller.
On-site need：Talk naturally with people in reception, guidance and delivery while protecting privacy on site.
What VXI adds：Local speech and language interaction; dialogue and video can stay on the device.
Early access is open now on our FPGA engineering evaluation platform; production specifications are defined together with early partners.
Map the robot model, controller platform, target model, and your priorities across power, latency and cost.
Import from open formats such as PyTorch and ONNX, then quantize and compile offline.
Run alongside your controller on the engineering evaluation platform to validate data flow and interfaces.
Measured accuracy, latency and power for your model, with reproducible steps.
Define production specifications, supply and rollout together, based on the results.
The engineers who build the silicon and toolchain work on your integration directly — no layers in between.
Every evaluation report lists test conditions and steps so your team can re-run them.
A long-term supply and version-compatibility plan, with details set out in the partnership agreement.
Evaluation scope, deliverables and commercial terms are confirmed in the partnership agreement for each project.
Tell us about your robot, controller platform and target model. Our engineers will plan the evaluation path with you.