Disengaged
No valid leader relationship. The follower holds rather than extrapolating motion from stale perception.
Connected autonomy · 2023–2024
Coordinate by
explicit state.
Can a group of small autonomous vehicles follow a leader, react coherently to a local obstacle, and resume without requiring a mechanically coupled platoon?
Auto-Platoon is a physical two-robot leader–follower system organized around “software latching”: explicit engage, follow, pause, disengage, and resume transitions shared through a networked coordination layer. The project integrates learned detection, image embeddings, monocular depth, feature tracking, Kalman estimation, obstacle sensing, and low-level motion.
Authors
PublicationarXiv:2405.11659 · May 2024
Physical scopeTwo Raspberry Pi mobile robots + host server
Research premise
A safe platoon is not merely a follower controller. It is a distributed agreement about when the group is engaged, when one agent’s observation should stop everyone, and when the formation may resume.
Two-robot physical validationFollow · obstacle-triggered pause · coordinated resume
Core idea
A latch sits between confidence and motion.
The perception stack proposes a relationship between leader and follower. The latch turns that continuous, uncertain estimate into an explicit coordination state. A robot may be visible without the platoon being engaged, and a cleared obstacle does not imply immediate motion until the re-engagement conditions are satisfied.
Latched
Identity, distance, and communication conditions support a leader–follower relationship; following commands are permitted.
Paused
An obstacle or peer-stop message suppresses motion across the platoon while preserving the coordination context.
Reacquiring
Perception and estimation re-establish the leader and safe spacing before motion resumes.
Resumed
The latch returns to an engaged state only after the relevant obstacle and tracking conditions clear.
Group behavior
The state machine converts a local observation into a coordinated response instead of relying on independent reactive controllers.
Architecture
Perceive, estimate, decide, communicate, control.
The stack uses multiple perception routes because no single signal is sufficient across all situations. Appearance maintains identity, features support tracking, monocular depth estimates spacing, and obstacle sensing can override the following objective.
01PerceiveYOLOv7 · feature tracking
02RepresentMediaPipe image embedding
03EstimateMiDaS depth · Kalman filter
04Coordinatesocket messages · latch state
05Controldistance hold · stop · resume
| Module | Purpose | Failure it addresses |
|---|---|---|
| Custom YOLOv7 | Detect the leader vehicle and relevant scene objects. | Initialization and reacquisition after tracking loss. |
| Feature tracking | Maintain the target between heavier detections. | Frame-to-frame identity continuity. |
| Image embedder | Compare appearance for leader identity. | Latching onto the wrong similar object. |
| MiDaS depth | Estimate relative separation from monocular imagery. | Following too closely or losing useful spacing. |
| Kalman filter | Regularize noisy position and motion estimates. | Control oscillation from frame-level noise. |
| Obstacle and peer messages | Propagate stop and resume state. | One follower continuing while another must stop. |
Architecture breakdown
The published diagrams expose the decisions between modules.
I designed the overall process architecture and the published subsystem architectures: the dynamic planner, cooperative sensing and communication topology, and close-range controller. The implementation and paper were collaborative; the contribution here is stated at the architecture level rather than implied by a generic system description.
Trace one module
01 · Perception and tracking
Establish identity before motion.
A custom YOLOv7 model initializes the leader. MobileNetV2 embeddings and cosine similarity associate it across frames, while a Kalman filter estimates centroid, scale, and motion. MiDaS supplies the relative depth used downstream.
- Dataset
- 4,200 train · 400 validation · 200 test
- Tracker
- appearance embedding + state estimation
- Range
- relative depth + reference calibration
02 · Dynamic planning
Choose follow, stop, or search.
Tracked target state and depth enter a planner that distinguishes a leader from an obstacle. A valid leader produces a follow plan; an unsafe depth produces stop-and-proceed behavior; target loss triggers leader search.
03 · Cooperative sensing
Turn a local event into group state.
The main server processes images and returns tracking outputs. Followers report status to the leader’s local server and read the current system state, allowing one agent’s obstacle event to stop the coordinated system.
04 · Close-range control
Align first, then regulate distance.
The controller combines target state, IMU orientation, and monocular depth. It corrects angular misalignment, evaluates the leader-distance threshold, then commands trace or stop behavior.
Intermediate evidence from the perception stack
The page now retains the artifacts that connect the architecture to the physical result: dataset construction, temporal tracking, depth estimation, and detector outputs.
Physical implementation
Coordination lives across three computational layers.
A host computer performs the heavier perception and state logic. Raspberry Pi computers on the leader and follower bridge network decisions to local sensors and motors.
Perception and coordination server
Python, OpenCV, learned models, socket communication, and state estimation run on the central machine.
Motion and group reference
The leader supplies the behavior that followers reproduce while also responding to coordinated stop messages.
Tracking and local safety
The follower estimates separation, commands its motors, detects dynamic obstacles, and publishes state changes back to the group.
- 01
Acquire and identify the leader
Detection, appearance, and temporal features establish the candidate leader before latching.
- 02
Estimate relative motion
Monocular depth and filtered image measurements approximate the state required by the following controller.
- 03
Maintain a threshold distance
The follower advances, holds, or slows as the estimated gap changes.
- 04
Propagate a safety event
A dynamic obstacle triggers a pause message so the leader and other followers do not continue independently.
- 05
Revalidate and resume
After the obstacle clears and the leader remains tracked, the latch permits coordinated motion again.
Evidence
A physical behavior result, not yet a benchmark.
The available artifact demonstrates the complete state sequence on two robots: leader acquisition, following, obstacle-triggered group stop, and automatic resume. The paper and repository establish integration; they do not report a sufficiently controlled repeated-trial matrix for statistical performance claims.
Coordination sequence
Extracting the state transition from the video makes the implementation visible: nominal spacing, obstacle introduction, group pause, obstacle clearance, and resumed following.
DemonstratedLeader tracking and distance-based following
Physical video
DemonstratedObstacle-triggered coordinated stop
Physical video
DemonstratedAutomatic resume after clearance
Physical video
Not measuredLatency, repeatability, false latch and formation error
Needs logs
Evidence needed for a stronger claim: repeated trials with timestamped leader pose, follower pose, latch state, obstacle state, network delay, stop reaction time, resume delay, and false engage/disengage events.
Contribution and limitations
System integration is the contribution.
The project is best read as a connected-autonomy prototype: a complete perception-to-coordination-to-control stack and a concrete mechanism for turning confidence into group state.
Software latching made uncertainty and coordination explicit.
I designed the end-to-end process architecture and each published subsystem architecture: the perception-to-planning flow, dynamic planner, cooperative sensing and communication network, and close-range controller. I also integrated the perception, estimation, communication, and controller behavior into the physical leader–follower demonstration. The implementation and paper were collaborative, and all co-authors are credited above.
Limitations. Validation uses two small robots and a host-centric architecture. Scale to longer platoons, packet loss, network partitions, outdoor conditions, heterogeneous robots, high speed, and formal string stability was not established.
Citation
Tharun V. Puthanveettil, Abhijay Singh, Yashveer Jain, Vinay Bukka, and Sameer Arjun S. “Auto-Platoon: Freight by Example.” arXiv:2405.11659, 2024.