Energy-Aware Learning for Optimized Drone Fleet Coordination in Urban Delivery

Lab: HiPeRLab, UC Berkeley

Timeline: 2025–2026

Status: Completed

Developed a data-driven energy prediction pipeline for quadrotor flight using real-world telemetry data from the HiPeR Lab at UC Berkeley. The goal was to enable energy-aware fleet coordination for urban drone delivery by accurately predicting energy consumption under varied flight conditions.

Python PyTorch Transformers LSTM Bi-LSTM RNN Sim-to-Real

Key Results

95%

Prediction accuracy (R²)

Architectures benchmarked

Sim→Real

Validation framework built

System architecture for energy-aware drone fleet coordination

Fig 1. End-to-end system architecture; from flight simulation through observation construction to energy prediction

Problem & Motivation

Urban drone delivery requires precise energy management - a drone that runs out of battery mid-flight is not just an inconvenience, it's a safety hazard. Existing energy models rely on simplified physics that don't capture the nonlinear dynamics of real quadrotor flight (wind disturbances, payload variation, battery degradation).

This project asked: can we learn energy consumption models directly from flight data that are accurate enough to enable fleet-level coordination and routing decisions?

Approach

Using real-world telemetry data collected from the HiPeR Lab's quadrotor testbed, I built and compared four deep learning architectures for energy prediction:

Vanilla RNN — baseline sequential model
LSTM — captures long-term dependencies in flight trajectories
Bi-LSTM — bidirectional context for improved temporal modeling
Transformer — attention-based architecture for global sequence understanding

Domain randomization and training pipeline

Fig 2. Domain randomization strategy and model training pipeline with architecture benchmarking

Each model was trained on flight telemetry features (velocity, acceleration, orientation, motor commands) to predict cumulative energy consumption. I designed a consistent evaluation framework to compare convergence, accuracy, and generalization across flight conditions.

Results

Training convergence comparison across GRU, MNN, and LSTM

Fig 3. Training convergence - GRU, MNN, and LSTM all converge to ~0.05 loss

LSTM predicted vs ground truth normalized power

Fig 4. LSTM prediction vs ground truth - R² = 0.95

The best-performing model achieved 95% prediction accuracy (R²), demonstrating that data-driven approaches can reliably model the complex energy dynamics of quadrotor flight. The prediction is temporally smooth and tracks major power trends. p95 forward latency < 0.2 ms/step, well under the 20 ms (50 Hz) control budget.

Sim-to-Real Validation

A key contribution was the development of a sim-to-real validation framework. Models trained on simulation data were evaluated against real-world flight telemetry to assess deployment feasibility and identify the domain gap.

[ To be updated soon: sim-to-real comparison plot (ongoing)]

Fig 5. Sim-to-real transfer analysis

What I Learned

This project deepened my understanding of sequence modeling for physical systems, the challenges of sim-to-real transfer, and the practical considerations of deploying ML models for safety-critical robotics applications. Working with real flight data, with all its noise and edge cases, was fundamentally different from clean simulation datasets.