Many angles. One story. Directed by AI.
Singularity is a three-person international team building intelligent systems at the intersection of AI, filmmaking, real-time media, and robotics.
Our main product is Dolly: a local-first autonomous production system that can see a scene, hear its creators, make directing decisions, control real production tools, assemble footage, and learn the creator's style.
We are developing Dolly for AdventureX 2026 in Hangzhou, China.
Traditional production requires a camera operator, director, technical director, editor, and production assistant. Dolly brings these roles into one coordinated AI system.
Dolly connects:
- Real multi-camera video
- Stereo microphone input
- Person detection and target tracking
- Active-speaker awareness
- Typed and voice commands
- AI-assisted directing
- OBS live production
- Unitree Go2 robot cinematography
- Local post-production with KinoCut
- Persistent project and creator-style memory
Dolly does not just recommend an action. It executes the action and requires real evidence before reporting success.
flowchart LR
A["Cameras · Mic · Go2"] --> B["Vision · Audio · Tracking"]
B --> C["DirectorPlan · Safety"]
C --> D["OBS · Robot · KinoCut"]
D --> E["Readback · Receipts · Memory"]
E -. "learn" .-> C
Every production action follows the same closed loop:
Observe → Understand → Plan → Validate → Execute → Verify → Learn
If a camera is stale, OBS does not confirm a switch, the robot loses its heartbeat, or an editor job does not return a valid output, Dolly reports the action as unavailable or failed. It never invents success.
| Pillar | What Dolly does |
|---|---|
| SEE | Multi-camera vision, person detection, target tracking, active-speaker awareness, robot vision, and view scanning |
| HEAR | Microphone input, audio activity detection, Whisper speech recognition, and natural voice commands |
| DIRECT | Typed and voice control, structured DirectorPlans, automatic camera selection, shot decisions, and production commands |
| EXECUTE | Real OBS switching, recording control, Unitree Go2 actions, and local KinoCut editing |
| LEARN | Creator shot preferences, pacing, accepted or rejected edits, Dolly Lens style memory, and Dolly Save project memory |
| PROVE | OBS readback, robot acknowledgements, render receipts, event timelines, and fail-safe system state |
Dolly receives real camera and audio signals, evaluates the scene, and selects the most useful shot. Creators can also control the system with direct commands:
"Dolly, action."
"Camera one."
"Follow this person."
"Dolly, again."
"Dolly, hold."
"Dolly, cut."
Simple and safety-critical commands remain deterministic and local. AI planning is schema-constrained and cannot issue arbitrary code, URLs, or unsafe motor commands.
For each action, Dolly checks the system that performed it:
- OBS scene changes require program-scene readback.
- Camera feeds require fresh, advancing frames.
- Robot actions require guarded commands and acknowledgements.
- Editing jobs require a real render receipt and output file.
- Failed compound actions stop the remaining steps.
After the creator says “Dolly, cut,” recorded media moves into a structured editing workflow:
Recordings → Ingest → Edit Plan → KinoCut → Draft → Review → Export
Source footage stays immutable. Dolly produces versioned derived outputs and preserves previous takes, feedback, and draft history.
Dolly Lens learns from real creative decisions:
- Preferred shot types and framing
- Editing pace and duration
- Accepted and rejected cuts
- Retakes, holds, and manual overrides
- Optional robot-shot preferences
Dolly Save preserves projects, sessions, camera roles, takes, commands, edit versions, feedback, and export state.
Dolly extends beyond fixed cameras through the Unitree Go2 Air and DimOS.
The robot system is designed to provide:
- A live robot-camera feed
- View scanning and target selection
- Person detection and continuous tracking
- Guarded follow-person behavior
- Heartbeat and stale-frame monitoring
- Immediate E-STOP and target-loss protection
- An OBS-compatible video bridge
The robot gateway, API contracts, command queue, MJPEG layer, state mapping, TTL protection, and software safety foundation are implemented. Real DimOS video, detection, follow, and physical-stop verification are the current integration focus.
Dolly keeps humans in control of the production.
- Persistent emergency stop
- Stop commands bypass AI planning and queued work
- Explicit local arming before physical motion
- Stale-frame and frozen-video detection
- Robot heartbeat monitoring
- Target-loss automatic stop
- Command allowlists and strict schemas
- Rate, speed, and range limits
- Duplicate and replay protection
- Honest degraded and unavailable states
| System | Status |
|---|---|
| Two-camera capture and live preview | Working |
| OBS authentication, switching, and readback | Working |
| Typed command and DirectorPlan pipeline | Working |
| Person detection and tracking pipeline | Implemented |
| Stereo audio and speech pipeline | Implemented; final hardware validation ongoing |
| KinoCut local editing and render receipts | Working |
| Project persistence and style memory | Implemented |
| Go2 gateway and software safety layer | Working |
| Real DimOS camera, tracking, and follow | Integration in progress |
| Full recording-to-draft demo | Final integration in progress |
The main Dolly production platform:
- FastAPI orchestration backend
- React and TypeScript production interface
- Camera and audio discovery
- Vision, tracking, and directing pipeline
- OBS WebSocket control and readback
- Whisper speech pipeline
- KinoCut post-production
- SQLite-backed projects and creator memory
- Safety, receipts, diagnostics, and automated tests
The robotic cinematography layer:
- Unitree Go2 and DimOS gateway
- Robot state and command contracts
- MJPEG video bridge
- Command queue and TTL controls
- Heartbeat and stale-frame safety
- Scan, target, follow, and stop interfaces
- Integration boundary for Dolly and OBS
Python · FastAPI · React · TypeScript · WebSockets · OBS Studio · OpenCV · YOLO · ByteTrack · Whisper · FFmpeg · KinoCut · SQLite · Unitree Go2 · DimOS
| Member | Role | Focus |
|---|---|---|
| Bayarbayasgalan Enkhtulga — Nero | Systems Architect & Integration Lead | Dolly architecture, backend and APIs, cameras, audio, OBS, command pipeline, production UI, post-production, and end-to-end integration |
| Ceaser Zhao | Robotics & AI Lead | Unitree Go2, DimOS, robot gateway, video bridge, perception, tracking, follow behavior, and physical safety |
| Jeboy | Product, UX & Pitch Lead | Product strategy, creator workflow, UX, demo direction, storytelling, presentation, and user validation |
- Local first — core production continues without depending on a cloud model.
- Evidence first — every successful action must have a receipt or readback.
- Creator controlled — automation assists the creator; it does not take ownership away.
- Fail safe — uncertainty, stale data, or lost connections produce a safe stop.
- Privacy conscious — media and editing stay local wherever possible.
- Built for reality — real devices, real production software, and honest system state.
We believe the future of production is not a single smarter camera. It is a coordinated intelligence that understands the whole creative process.
Dolly is our first step toward a production crew that can see, hear, direct, execute, verify, and learn — while keeping the creator in control.
