Source-linked AI summary
Communication Reduction via Semantic-Based Encoding in DMPC Using LSTMs
Torben Schiz, Pedro H. J. Nardelli, Henrik Ebel
TL;DR
DMPC can overwhelm wireless communication because agents exchange substantial information every time step. The paper uses LSTM-based semantic encoder-decoder networks to reduce messages, achieving performance close to full communication and resilience when full communication falters.
Problem
DMPC requires substantial inter-agent communication, while real-world robotic applications remain limited and semantic reduction needs evaluation for networked control.
Method
The paper trains LSTM-based encoder-decoder networks that transmit latent message representations and reconstruct control information within distributed optimization.
Results
SHRED closely matches full-communication convergence in simulation and performs well with physical wireless communication where full communication falters.
Takeaways & Limitations
LSTM-based semantic communication can reduce DMPC communication while retaining convergence under some conditions that overwhelm unreduced communication.
Takeaways & Limitations
LSTM decoders add computational overhead that grows with formation size and prediction horizon, potentially shifting the bottleneck from communication to computation.
Abstract
from arXiv · showhide
The communication demands of distributed model prediction control (DMPC) can overwhelm even advanced wireless communication technologies as agents must exchange a significant amount of information at least once per time step. To semantically reduce communication demands, this work employs encoder-decoder networks built around long-short term memory (LSTM) cells in a distributed optimization algorithm. Agents publish a reduced representation of a message and receivers reconstruct the original message upon reception. In tests with reduced communication using formations of mobile robots, trained networks retain satisfactory performance and work reliably under conditions overwhelming full communication. As the results show, the usage of LSTMs either allows unprecedented reconstruction accuracy or the usage of different prediction-horizon lengths without the necessity to retrain.
1 Introduction
The introduction identifies communication as a major barrier to real-world DMPC applications, especially in robotics, and motivates semantic encoding to reduce network traffic while preserving control performance. The article proposes machine-learning-based latent-space compression at the application level, retaining compatibility with standard communication technologies and addressing communication-induced sim-to-real challenges.
- Problem: Real-world DMPC applications are presently lacking, particularly in robotics, despite the field’s worthwhile networked-control applications.The authors identify this gap as a central motivation for their work.
- Motivation: Semantic communications represent and transmit task-relevant information using fewer communication resources, potentially reducing network traffic and facilitating timely delivery.The approach may increase computational complexity at transmitters and receivers.
- Contribution: The article proposes application-level, machine-learning-based reduction of communicated data into a latent space without changing lower communication-stack elements.This preserves straightforward applicability with standard communication technologies and protocols.
- Limitation: Communication introduces a significant additional sim-to-real gap in DMPC for mobile robots.This gap is distinct from the usual complexity-driven simulation-to-reality gap associated with real robot dynamics.
- Related Work: Prior data-driven work reduced inter-agent communication with an autoencoder that encoded messages before transmission and decoded them after reception.Simulated robot formations successfully completed tasks with this approach.
2 Problem Setting
The problem setting formulates distributed receding-horizon MPC for dynamically decoupled agents coupled through a common cost, then instantiates it as formation control for nonholonomic mobile robots. Each robot optimizes locally while exchanging predicted inputs with neighbors through distributed optimization.
- General distributed MPC: DMPC couples N ≥2 dynamically decoupled systems through a common finite-horizon cost and solves the resulting optimal-control problem in receding-horizon fashion.The prediction horizon has length H, with states z and inputs u subject to independent per-system constraints.
- General distributed MPC: The multi-agent dynamics are formed by concatenating individually decoupled robot dynamics, with each predicted trajectory governed by z(t + k + 1 | t) = f(z(t + k | t), u(t + k | t)).The notation (· | t) denotes trajectories predicted at time t across the prediction horizon.
- Formation-control application: The test system is a formation of nonholonomic differential-drive robots whose Pfaffian constraint prevents instantaneous lateral motion.Each robot is modeled in a two-dimensional plane using position, orientation, linear velocity, and angular velocity.
- Formation-control application: Formation outputs describe robot geometry relative to the geometric center in an auxiliary reference frame rotated by the desired formation orientation.The output is expressed as yR(z, ˆθd) = CR(ˆθd)z.
- Distributed optimization: At every time step, each robot line-searches a candidate sequence for its own inputs, exchanges that prediction with neighboring systems, and iterates toward a convex-like solution.Each robot optimizes only its own control inputs, while inputs obey box constraints.
3 Neural Network Architectures for Communication Reduction
The section presents five LSTM-based encoder-decoder architectures for reducing inter-agent communication in DMPC. Each agent compresses its predicted input sequence into a latent representation, communicates it, and reconstructs the sequence with a decoder; the architectures differ in decoder type and communicated states.
- Architecture overview: Five architectures share an LSTM-based encoder, while SHRED uniquely uses a feed-forward decoder instead of an LSTM decoder.The architectures are motivated by limited encoding/decoding time and restricted data transfer from encoder to decoder.
- Architecture overview: Encoder-decoder LSTMs support multiple prediction-horizon lengths from a single training run, whereas SHRED requires retraining when the horizon changes.LSTM decoders reconstruct predictions in reverse order, while SHRED communicates the encoder output as decoder input.
- Semantic communication: Each agent compresses its predicted input sequence into a latent representation before publication, and receiving robots reconstruct it with a decoder.The same pre-trained model is used across robots, and messages are reduced and reconstructed individually, making communication reduction independent of formation size.
- LSTM variants: Cell LSTM communicates only the encoder cell state, initializes decoder input and hidden state to zero, and uses the received cell state as the decoder’s initial cell state.This architecture is illustrated as one of the considered LSTM variants.
- LSTM variants: LSTMP communicates the cell state and projected hidden state, reducing LSTM output dimensionality to that of a velocity pair through projection layers.The projected hidden state initializes the decoder input, while the encoder cell state initializes the decoder cell state and decoder hidden states start at zero.
- LSTM variants: LSTM+FFL projects only the encoder output to a two-dimensional velocity-pair representation, communicates it with the cell state, and uses both to initialize the decoder.LSTM+FFLH instead communicates only the encoder output, uses it as the decoder’s initial hidden state and projected initial input, and initializes the cell state to zero.
4 Data Generation, Tuning, and Training
The study generates communication data from two-robot formations, tunes encoder-decoder architectures, and trains models for fixed or variable prediction horizons. Training uses standardized dataset splits, optimization settings, and a 10-dimensional latent code, with SHRED achieving the best validation performance among the variants.
- Workflow: The workflow comprises data generation from inter-agent communication, standalone encoder-decoder training and tuning, and integration of the model components into the optimization algorithm.These stages define the learning-based communication strategy for DMPC.
- Dataset: Two-robot formations provide the data because, without disturbances, robots in a formation make similar predictions and cover the input space similarly well.Only the communicated predictions of one robot are used for data generation.
- Dataset: 2000 randomly generated scenarios train fixed-horizon models at H = 20, while 200 scenarios per H ∈ Z21:25 support variable-horizon LSTM training.The variable-horizon scenarios are a random subset of the 2000 scenarios used for H = 20.
- Training setup: The dataset is split into 90% training data and 10% validation data, with hyperparameters tuned for each neural-network architecture.Models are implemented in PyTorch, and training runs are tracked using Weights & Biases.
- Training setup: 200 epochs is the upper limit for tuning runs, with batch sizes 265, 512, and 1024, ADAM optimization, mean-square error loss, and code size 10.Learning has mostly ceased after 200 epochs, motivating the selected upper limit.
- Architecture comparison: The SHRED-based variant yields the best validation loss, whereas the LSTMP variant yields the worst training performance among the considered architectures.For LSTM+FFL, hyperbolic tangent outperforms ReLU in testing despite ReLU working slightly better during training.
5 Test: Simulative Analysis
The simulative analysis tests convergence and communication-reduction methods under idealized, lossless communication using formations of two to six robots. LSTM-based architectures achieve the lowest average cost across most trained prediction horizons and retain this behavior outside the training range, while horizon-specific training improves accuracy on that horizon.
- Test setup: The test evaluates convergence with trained models and reduced communication while all processes run on one computer without packet loss.The subsequent section addresses performance under demanding, realistic wireless conditions.
- Test setup: 200 scenarios test formation sizes from two to six robots with H = 20, using Tend = 200 s, 300 s, or 450 s depending on formation size.Larger formations receive longer simulations to permit empirical convergence testing.
- Prediction-horizon analysis: Methods trained for variable horizon lengths perform worse on a specific horizon than methods trained specifically for that horizon because they trade accuracy for generality.The H = 20 findings are also supported by the value-function plot for one example scenario.
- Error analysis: CDFs of y-error at Tend = 300 s compare full and reduced communication across 60 formation scenarios involving two to four robots.The CDFs confirm the preceding observations, with full communication achieving the lowest y-error.
- Prediction-horizon analysis: Architectures combining an LSTM element with an FFL projection have the lowest average cost at Tend for most horizons covered during training.The same behavior is observed for H = 18, which lies outside the training range.
6 Numerical Experiments on Embedded Hardware
Numerical experiments on Raspberry Pi 5 hardware compared full and semantically reduced communication over 5 GHz Wi‑Fi. Reduced communication remained effective when full communication failed to deliver messages in time, although performance depended on the reduction method and experiment conditions.
- Experimental setup: The distributed optimization algorithm ran on dedicated Raspberry Pi 5 computers with 8 GB RAM, communicating through 5 GHz Wi‑Fi via UDP multicast.A personal computer connected by LAN simulated the mechanical dynamics.
- Baseline experiment: 99% of communicated information was received by all three communication configurations for eight robots with H = 20 and p̄ = 3.The configurations also displayed convergence similar to the simulation results.
- Communication stress tests: At p̄ = 27 with four robots, full communication delivered no inter-agent messages in time, whereas reduced communication still converged to small value-function values.LSTM+FFLH received around 8 percentage points less information on average than SHRED, despite identical packet sizes and communicated information budgets.
- Communication stress tests: With eight agents, full communication collapsed at p̄ = 10, while LSTM+FFLH also struggled to converge despite receiving over 90% of information.Without communication reduction, the formation sometimes converged while receiving less than 40% of the communicated information.
7 Conclusion
The work proposes LSTM-based semantic encoding to reduce inter-agent communication in DMPC, with SHRED closely matching full-communication convergence in simulation and performing well under challenging physical wireless conditions. LSTM decoders support one model across multiple prediction horizons, but introduce computational overhead and do not match SHRED’s performance.
- The work proposes neural-network architectures based on LSTMs for semantic encoding to reduce inter-agent communication in distributed model predictive control.
- SHRED closely matches full communication’s convergence performance in simulation and performs well with physical wireless communication when full communication falters.
- SHRED outperforms the autoencoder-based communication-reduction method from.
- An LSTM decoder enables one model to support multiple prediction horizons, although no such model matches SHRED and decoder overhead increases with formation size or prediction-horizon length.
- An encoder-decoder LSTM converges under some conditions where unreduced communication is overwhelmed.