✅ This piece is a high-confidence 'inventory of what we own.' Its content comes entirely from the company's own exhibit materials (two PDFs, "IVIS Exhibit Introduction" and "Same, Vol. 2 (Speech Recognition)"), not from external research or speculation. Before judging "which way to go and whether we can make money," first spell out clearly "what exactly we hold in our hands"—otherwise it's easy to chase someone else's robot-dog / humanoid hype and drop the cards we actually have.
🌐 In one sentence: our card is not the "robot dog," it's the edge-AI layer called Reservoir Computing
Many discussions immediately get bogged down in "should we buy a Unitree or a Boston Dynamics robot dog?" But after taking inventory of the exhibits, you'll find: the common foundation of all four company demos is one and the same algorithm—Reservoir Computing (RC; Japanese "リザバー计算"; typical implementation: Echo State Network / ESN). The robot dog is merely one "demonstration vehicle" for it.
The characteristics of this RC capability layer are exactly what large models / heavy deep learning cannot do:
- Extremely low training cost: training requires only a single linear regression (least squares / ridge regression), unlike backpropagation that needs hundreds of epochs;
- Online incremental learning: using recursive least squares (RLS) to learn on the fly—adding a new category / a new person does not require retraining the whole model;
- Fits into small edge devices: already running inference + learning on FPGA (Kria KV260), Raspberry Pi 5, and pure CPU, at low power;
- Naturally good at prediction and anomaly detection of time-series dynamic patterns: precisely the safety-perception capability that "Physical AI / embodiment" most lacks.
This layer is the moat; the robot dog is just a booth prop. Below we dissect the four demos one by one, then return to "what this means for where we should go."
Source: all four demos were physical exhibits by the company at NexTech Week 2026 Spring (Physical AI–related exhibition area)1. Part of the Reservoir Computing "additional learning" results are the product of a NEDO (New Energy and Industrial Technology Development Organization) commissioned project (see the note under Demo ④).
🐕 Demo ① Quadruped Robot "Contactless Sensing" Collision Detection — Anomaly Detection with RC
What problem it solves. For a quadruped robot to roam safely in real environments, it must be able to detect contacts—such as "hitting a wall" or "getting snagged"—that obstruct normal walking. But installing contact sensors on each violently moving leg / part is very costly. This demo achieves "detecting collisions without installing contact sensors" using RC.
How it's done (core idea).
- Using only collision-free, contact-free normal walking data, train a model: "from the previous step's state variables of each leg + the control input (velocity), predict the current step's state variables";
- During normal walking, predicted ≈ measured;
- The moment a collision / contact occurs, the motion pattern is disrupted, and the deviation between predicted and measured values suddenly grows—this deviation is used to detect collisions. This is a classic case of "using RC's dynamic-pattern prediction capability to do anomaly detection."
Engineering details.
- Model structure: input layer of 24 dimensions (4 legs × previous-step joint-angle state variables 3-dim + control input 3-dim) → Reservoir layer → output layer predicting the current step's state;
- Output form: each leg's collision-detection value is published as a ROS2 message; in the GUI, collision intensity is shown by the size of a red circle (the circle briefly enlarges on collision; when snagged, the circle stays small but persists).
Why it matters (transfer value). The exhibit panel itself points out: this "RC anomaly detection" can transfer to a large variety of time-series data—speech / ambient sound, video, actuator control/state variables, sensor time-series of temperature·vibration·current·pressure, and network traffic. In other words, the true asset of this demo is not "robot-dog collision detection" but a reusable, lightweight time-series anomaly-detection engine—precisely the technical prototype of the later "predictive maintenance / equipment anomaly detection" business opportunity.
🔁 Demo ② Reservoir Computing for Efficient Control-Law Learning — Rotary Inverted Pendulum
What it demonstrates. Applying RC to control: inferring on an edge device to swing a rotary inverted pendulum (a motor rotates the arm horizontally to make the pendulum at the arm's tip stand upright) up from hanging down and stabilize it upright. The core selling point is "efficient reinforcement learning using a lightweight Reservoir-based 'world model'"—achieving the upright behavior with about 15 minutes of training on a GPU-equipped PC.
How the system is built (three parallel tracks).
- Real-machine trial-and-error (data collection): controlling the real machine while collecting world-model training data;
- World-model learning: using the Reservoir state to predict the next-moment angle, serving as a "simulator" of the controlled object;
- Policy learning (reinforcement learning): using PPO (Proximal Policy Optimization), leveraging the world model to run parallel simulations of the controlled object for efficient sampling.
Technical highlights.
- This is a POMDP task—the controlled object can only observe angle information (partially observable), making it harder;
- The world model and reinforcement learning share the same Reservoir (ESN), with an MLP readout layer—an extremely simple, computationally light structure;
- The outlook explicitly states the direction: adapting RC onto other methods for acceleration, autonomously adapting to environments on edge devices, applied to robot hand control and autonomous mobile robots.
Why it matters. It proves RC is not just "detection/classification" but can also enter the control loop—and as "sample-efficient, edge-self-adaptive" control at that—which opens up imaginative space for low-cost robots and cheap adaptive control of robotic arms/hands.
✋ Demo ③ RC Gesture Recognition on Edge FPGA — Ultra-Fast "Additional Learning"
What it demonstrates. Collecting data + learning on an edge device, recognizing dynamic image patterns (gestures) in real time; swapping in a different training set lets the same RC model recognize different patterns; and it implements an "ultra-fast additional learning" function.
System composition (all on the edge).
- Hardware: Kria KV260 FPGA board;
- Image feature extraction: using pretrained EfficientNet-Lite4 for dimensionality reduction;
- RC model: ESN, 2000 nodes;
- Learning method: recursive least squares (RLS)—it can perform incremental learning by "adding a new dataset n" on top of an already-learned dataset N, rather than, like backpropagation, feeding the entire dataset as one epoch repeatedly for hundreds of rounds.
- "Image feature extraction + Reservoir + RLS learning" is all implemented on the FPGA, achieving low power.
Why it matters. It puts both of RC's core advantages on the stage at once: ① true edge (FPGA low power) ② true online incremental learning (RLS). For scenarios that demand "fast on-site teaching, no network, no cloud, privacy-sensitive" (production lines, access control, beside equipment), this is a product path completely distinct from the mainstream "cloud-based large model."
🎙️ Demo ④ Edge Speaker Recognition — Pure-CPU Online Learning (Includes NEDO Results)
What it demonstrates. Collecting data + learning on an edge device, recognizing the speech patterns of multiple people in real time, and learning incrementally (online) in real time.
System composition.
- Device: Raspberry Pi 5 (inference + learning completed on CPU only);
- Speech features: Mel-Frequency Cepstral Coefficients (MFCC) and their first-order delta, second-order acceleration (10ms/step);
- RC model: ESN, 500 nodes;
- Learning method: RLS (recursive least squares);
- At inference time, silero-vad is used to segment out speech intervals, classifying and outputting only on valid speech segments.
Application directions pointed out (very important). The exhibit panel itself lists extended applications: ① speech-based detection of speaker physical-condition anomalies (health monitoring) ② anti-theft / security technology based on footstep-sound recognition. Neither of these is a "voice assistant"; both are "using sound time-series for anomaly / identity detection"—of one piece with the idea behind Demo ①.
📌 Important endorsement: The exhibit panel notes—"Part of the results related to the Reservoir Computing model's 'additional learning' are the outcome of a NEDO (New Energy and Industrial Technology Development Organization) commissioned project." This means the team already has a NEDO commission track record—a hard asset for later applications for national funding and for building external credibility (especially toward investors such as the chairman/boss), and it must be written into the business plan.
🧩 Converging the Four Demos into "One Capability"
| Demo | Vehicle/Hardware | RC Model | Learning Method | Reusable Capability Exposed |
|---|---|---|---|---|
| ① Sensorless collision detection | Quadruped robot + ROS2 | Reservoir (24-dim input) | Train on normal data, detect deviation | Time-series anomaly-detection engine |
| ② Rotary inverted pendulum control | Edge device + GPU training | ESN (shared by world model + RL) | PPO + world-model parallel simulation | Sample-efficient edge-adaptive control |
| ③ Gesture recognition | Kria KV260 FPGA | ESN 2000 nodes | RLS incremental learning | True edge + online incremental learning |
| ④ Speaker recognition | Raspberry Pi 5 (pure CPU) | ESN 500 nodes | RLS online learning | Identity / anomaly detection from sound time-series |
The four demos all point to one sentence—what the team truly commands is:
"An edge AI (Reservoir Computing) that can run on small devices such as MCU/FPGA/Pi, has extremely low training cost, can learn incrementally online, and is especially good at prediction and anomaly detection of time-series dynamic patterns—and that can connect to robots (ROS2) and into control loops."
🎯 What This Means for "Where to Make Money" (Leading into the Next Piece)
- Don't define ourselves as a "robot-dog company"—the hardware body (quadruped/humanoid) is capital-heavy, thin-margin, and brutally competitive; a 10-person team can't outcompete the full-machine makers;
- Define ourselves as the "perception / anomaly-detection / lightweight-control" software + module layer of embodied AI—this layer sits right on top of our RC moat: capital-light, licensable, suitable for commissioned work, and able to grow into products;
- The four demos already naturally point to several directions with real paying customers: equipment predictive maintenance / anomaly detection, robot safety (sensorless collision detection packaged as a licensable module), and sound × health/security monitoring.
Whether these directions are "worth doing, who will pay, and whether they can survive 3 years" will be checked point by point with external market intelligence in the next piece, Embodied AI Directional Research.