Distributed Swarm Algorithm

SWARMBENCH-30 (DTAR-Lite) Submission

This repository contains our submission for the Swavlamban Innovathon 2025 – Development of Distributed Swarm Algorithm, evaluated using the SWARMBENCH-30 (DTAR-Lite) benchmark.

Our solution implements a fully decentralized, single-file swarm agent for dynamic task allocation under lossy radio communication, without any central controller, shared memory, or synchronization primitives.

---

1. Problem Overview

SWARMBENCH-30 simulates a constrained multi-agent maritime environment with:

• Dynamically appearing spatial tasks
• Hard deadlines and service times
• Heterogeneous task values
• Multiple mobile agents with speed limits
• Lossy, bandwidth-limited broadcast communication

Key Constraints

• Deterministic execution (fixed seeds)
• CPU-only, single-threaded
• All-to-all broadcast with packet loss
• ≥95% stable task ownership within 60s
• Penalization for deadline violations or non-convergence

Scoring Function

score = 0.6 × value_ratio − 0.2 × norm_distance − 0.2 × norm_bytes

---

2. Algorithm Summary

We implement a distributed auction-based task allocation algorithm with the following properties:

• No central planner or coordinator
• Local task evaluation by each agent
• Global consensus via repeated claim broadcasts
• Conservative feasibility filtering to avoid penalties
• Capability-aware commitment for constrained tasks (S7)
• Lightweight leader heartbeat for fault detection only (S6)

Task allocation emerges from peer-to-peer interaction, not centralized decisions.

---

3. Agent Architecture

Each agent executes identical logic and maintains only local state plus information learned from peer messages.

Core State

• Agent ID and maximum speed
• Fixed capability (thermal, lift1, sea3)
• Local task bundle (my_tasks)
• Global bid table learned from peers (world_bids)
• Current executing task
• Completed task history

Communication Model

• Broadcast-only messages

• No acknowledgments or retries

• Two message types:

  • claim → task ownership
  • role → leader heartbeat (S6)

• Messages are sent only on state change to minimize bandwidth

---

4. Functional Breakdown (Code-Mapped)

Distance Utility

dist(ax, ay, bx, by) Computes Euclidean distance. Used for bidding, reachability, and navigation.

---

Capability Gate (S7)

can_perform_task(task)

• Enforces hard capability constraints
• Tasks requiring a capability are invisible to non-matching agents
• Prevents invalid bidding at the source

---

Deadline Feasibility

task_is_reachable(task, t, x, y)

Ensures a task can be completed safely before its deadline:

travel_time = distance / speed
total_time  = travel_time + service_time
reachable   = total_time < 0.93 × (deadline − current_time)

The 7% safety margin prevents deadline violations caused by motion discretization and communication delays.

---

Bid Computation (Core Logic)

calculate_task_bid(...)

Each feasible task is scored using a multiplicative utility model:

bid =
value_component
× distance_component
× urgency_component
× capability_multiplier
× load_factor
× spatial_factor
× sequence_factor

Key design choices:

• Exponential value preference (high-value dominance)
• Inverse distance penalty (fuel efficiency)
• Exponential urgency near deadlines
• Strong capability prioritization (S7)
• Load balancing to avoid over-commitment
• Spatial bias to reduce global travel
• Task chaining to prevent zig-zag motion

Infeasible tasks return a large negative bid.

---

Bundle Construction

build_task_bundle(...)

CBBA-inspired, two-phase selection:

1. Capability tasks first (S7)

   • Select highest-priority matching task
   • Lock agent (cap_mode) until completion

2. Normal tasks

   • Select up to 1–2 tasks
   • Skip tasks dominated by other agents’ bids

Small bundles improve convergence speed and deadline safety.

---

Distributed Consensus

update_from_messages(...)

• Processes incoming claim messages
• Updates global bid table
• Drops tasks when clearly outbid
• Tracks leader heartbeats (S6)

Consensus emerges via repeated broadcasts under lossy conditions.

---

Leader Failure Detection (S6)

handle_leader_election(...)

• Leader sends heartbeat every 0.5s
• Followers timeout after 1.0s
• Any agent may self-promote
• Higher term always wins

The leader does not assign tasks — task allocation remains decentralized.

---

Target Selection

select_target_task(...)

• Chooses which owned task to execute next

• Priority:

  1. Capability tasks
  2. Highest bid value

• Executes only tasks still owned by consensus

---

Navigation Control

navigate_to_target(...)

• Velocity vector toward target
• Stop threshold at 3.5 m (prevents oscillation)
• Linear deceleration within 20 m (smooth arrival)

Improves distance efficiency without affecting value completion.

---

Outgoing Communication

create_messages(...)

• Leader heartbeats at fixed rate
• Claim broadcasts only when bundle changes

Keeps bandwidth well below scenario limits.

---

Main Control Loop

step(t, dt, self_state, tasks_visible, inbox)

Executed at 10 Hz with strict ordering:

1. Detect task completion
2. Process incoming messages
3. Handle leader election
4. Cleanup stale data
5. Build task bundle
6. Select target
7. Compute motion
8. Generate messages

Returns:

{"vx": vx, "vy": vy}, messages

---

5. Performance Summary

Scenario	Result
S1–S6	Full value, zero penalties
S6	Instant leader recovery
S7	13 / 14 capability tasks
Packet Loss	Stable up to 20%
Convergence	~59s (within gate)

Average Score (S1–S7): 0.5664

---

6. Algorithmic References (Conceptual)

• CBBA — Distributed auction & bundle consensus principles Choi et al., IEEE Transactions on Robotics, 2009

• Raft — Heartbeat-based failure detection concept Ongaro & Ousterhout, USENIX ATC, 2014

These works inspired design choices; no direct implementations are used.

---

7. How to Run

Run all scenarios:

python run_all.py --team teams/nerdos_agent.py

8. Key Properties

• Fully deterministic
• Single-file agent
• No external dependencies
• No central coordination
• Designed explicitly for SWARMBENCH-30 constraints

---