Desktop Robot Lamp's Two-Month Evolution: Reads Emotions, Senses the Environment, and Supports Hot-Swappable AI Brains

The open-source desktop robot project Lamp made its debut two months ago as a purely experimental device — a 5-DOF robotic arm running STS3215 bus servos, powered by a single 12V cable, with a USB DAC to eliminate ground noise. Two months later, the project has delivered a substantial round of upgrades. Lamp has evolved from a quirky desktop gadget into an interactive companion capable of expressing emotions, perceiving its environment, and swapping out its "brain" on the fly.

From Robotic Readouts to an Emotionally Expressive Voice

Early Lamp used the kind of flat, monotone TTS voice that sounds like it's reading a status report. The most immediately noticeable change in this update is the switch to ElevenLabs v4 Turbo, which transforms the experience entirely: pitch and rhythm rise and fall naturally, pauses land where they should, and a joke actually sounds like a joke rather than a system notification.

Even more interesting is the addition of an emotional dimension. The developer deployed emotion2vec — a speech emotion recognition model — directly on-device to listen for emotional cues in the user's voice, then subtly reflect Lamp's current "mood" through the ring light on its head. This dual-channel feedback combining visual and audio cues makes interactions feel much closer to the natural emotional flow of real conversation.

What truly showcases the design's thoughtfulness is how personality is implemented. Lamp's character isn't hardcoded anywhere — it's a simple Markdown file that the agent parses at the start of each conversation. By default, Lamp presents itself as a warm, curious desktop companion that never breaks character to say things like "servo" or "API call." Change a few bullet points in that file, and you have a roommate with an entirely different personality. As the developer put it: you're editing the soul, not the code.

About emotion2vec: emotion2vec is an open-source speech emotion recognition model released by Alibaba DAMO Academy. Trained on a self-supervised learning framework, it extracts emotion embedding vectors directly from raw audio waveforms and supports recognizing multiple emotional categories including anger, happiness, sadness, and neutral. Its inference latency is low enough for real-time on-device use. Unlike traditional approaches based on keyword spotting or text-based sentiment analysis, emotion2vec processes acoustic features directly — meaning it can capture emotional signals from tone, speech rate, and pitch variation even when the user's words seem neutral. Deploying it on-device rather than in the cloud also eliminates network round-trip latency and protects the privacy of users' voice data.

It Can Actually "Read the Room"

In the first version, Lamp's only inputs were a camera and a microphone. This update adds a Sensirion SEN63C sensor over I2C — a compact integrated module that simultaneously reports particulate matter (PM1, PM2.5, PM4, PM10), temperature, humidity, and true CO2 concentration.

The data handling logic is worth noting from an architectural standpoint. The hardware abstraction layer (HAL) caches the latest environmental snapshot in RAM, which is polled periodically by a background worker in the OS. Lamp doesn't stuff every raw sample into the conversation context — it only fires an event when an environmental metric shows a trending drift. This "only speak when there's something worth saying" design avoids meaningless information overload. The data is also accessible in read-only form via Web and MQTT.

The real-world result is refreshingly practical: when someone was cooking downstairs and the developer closed the window, Lamp flagged that the room was getting stuffy before he noticed it himself. Since the agent layer only queries abstract key-value pairs (pm, co2, temperature, etc.), it's completely agnostic to whatever sensor hardware is actually soldered underneath — leaving plenty of room to swap or expand sensors down the line.

About Sensirion SEN63C: The SEN63C is an all-in-one environmental sensor module that integrates laser particle counting, NDIR (Non-Dispersive Infrared) CO2 detection, and temperature/humidity measurement. The CO2 detection uses the NDIR principle, calculating concentration by measuring the absorption of infrared light at a specific wavelength as it passes through the gas. Compared to older MOX (metal oxide) equivalent CO2 (eCO2) sensors, NDIR measures the actual concentration of CO2 molecules — making the data more accurate and reliable, without false readings triggered by other volatile organic compounds in the air. This distinction matters greatly in indoor air quality monitoring: elevated eCO2 readings are often just interference from cooking fumes or cleaning products, not a genuine ventilation problem. The addition of MQTT also allows sensor data to be subscribed to by smart home platforms (like Home Assistant) or custom scripts in a lightweight publish/subscribe fashion, further extending how the data can be used.

A Hot-Swappable "Brain"

The most architecturally significant improvement is the modularization of the agent backend. Early Lamp was locked into a single agent runtime. Now, the Autonomous OS treats the agent as a pluggable backend defined in a configuration file, fully decoupled from personality and long-term memory.

The developer demonstrated this with a compelling example: by changing a single config entry, he migrated Lamp from Hermes to OpenClaw^Ad. Throughout the process, Lamp retained the exact same voice, the same quirks, and everything it had previously learned about its owner. The agent handles reasoning, while the soul and memory survive the "brain transplant" completely intact.

This decoupling philosophy has broad relevance for open-source robotics projects: when the inference model, personality definition, and long-term memory are all independent of each other, rapid iteration on the underlying model doesn't force you to start from scratch. You can switch to a more capable agent backend at any time without losing continuity in the user relationship.

About Hermes and OpenClaw: Both Hermes and OpenClaw belong to the local LLM inference runtime ecosystem. The Hermes series consists of instruction-tuned open-source models released by Nous Research, known for strong function calling and role-playing capabilities — commonly used in agentic tasks requiring structured output. OpenClaw is a runtime framework optimized for tool-calling scenarios, emphasizing low latency and more flexible toolchain integration. The essence of "hot-swappability" is that Autonomous OS maintains a standardized interface between the agent backend and the upper-layer services: the personality file, long-term memory database, and tool registry are all stored in independent formats that any compliant inference backend can read from and write to directly, enabling seamless "brain" swaps. This architecture is especially practical in the fast-moving open-source model ecosystem — when the community produces a better base model every few weeks, you can continuously enjoy the latest reasoning capabilities without rebuilding your accumulated personality and memory.

Why These Updates Matter

Lamp's three upgrades are each independently useful, but together they point toward a more mature product philosophy: decoupling "hardware, perception, personality, and reasoning" into separate layers so each can evolve on its own. Emotionally expressive voice solves the warmth problem in interaction; environmental sensing solves the contextual awareness problem; and the hot-swappable agent backend solves the long-term maintainability problem for the technology stack.

For hobbyists who want to build their own desktop AI companion, this configuration-and-Markdown-file-driven architecture dramatically lowers the barrier to customization — you don't need to be a programmer to reshape its personality or swap out its sensing hardware. That's precisely what makes open-source hardware projects so compelling: the boundaries of the product are defined by the community and its users together.