TL;DR
BoBDeLaN (Bounded-Bandwidth Deep Lagrangian Networks) introduces a robust framework for adaptive robot control by replacing standard LSTM encoders with gain-bounded Recurrent Equilibrium Networks (RENs). By combining these certified recurrent models with frequency-shaped filtering, the authors provide mathematical guarantees on disturbance sensitivity, resulting in significantly smoother and more reliable robot motion under varying payloads and sensor noise.
What Problem Does This Paper Address?
Deep Lagrangian Networks (DeLaN) are powerful because they embed the laws of physics—specifically Lagrangian mechanics—directly into neural networks. This ensures that learned robot models are energy-consistent and physically plausible. However, real-world robots often face "unmodeled dynamics," such as picking up an unknown heavy object or dealing with friction.
To handle this, researchers previously developed "Context-Aware DeLaN," which uses a history of measurements to infer a "context" vector that adjusts the model on the fly. Usually, an LSTM (Long Short-Term Memory) network acts as the encoder. The problem? LSTMs are black boxes when it comes to stability. They can be overly sensitive to high-frequency sensor noise, leading to erratic context updates that cause the robot to jitter or become unstable. There are no inherent guarantees that an LSTM will "forget" its initial state or stay within safe sensitivity bounds when faced with disturbances.
Methods
The core innovation of BoBDeLaN is the integration of stability-by-design into the context adaptation loop.
- REN Context Encoder: The authors replace the LSTM with a Recurrent Equilibrium Network (REN). Unlike standard RNNs, RENs are parameterized to satisfy specific mathematical properties by construction. They guarantee contraction (the network exponentially forgets its initial state) and a bounded incremental $\ell_2$ gain (the output's change is strictly bounded relative to the input's change).
- Frequency-Shaped Sensitivity: To specifically target payload adaptation, the authors precede the REN with a cascaded low-pass filter. This creates a "frequency-weighted" sensitivity: the model remains responsive to slow, meaningful changes (like a new payload) but becomes "deaf" to high-frequency measurement noise.
- Theoretical Bounds: The paper derives finite-horizon bounds that link the encoder's gain to the final residual torque predictions. This allows researchers to calculate a "safety envelope" for how much a sensor disturbance can possibly affect the robot's control signal.
- Control Integration: The architecture is integrated into a Model Predictive Control (MPC) framework (CaDeLaC), as shown in the system diagram below.
Datasets
The authors evaluated BoBDeLaN using both simulated and real-world data:
- Simulation: A dataset generated in MuJoCo featuring 2,020 trajectories of a robot arm with 100 different payload configurations (masses ranging from 0 to 4 kg).
- Real-Robot Data: Recorded trajectories from a Franka Panda robot performing manipulation tasks (e.g., drawing an infinity symbol) with payloads of 1, 2, and 3 kg.
- Hardware Benchmark: A real-world pick-and-place task involving dumbbells of different weights to test rapid adaptation.
Key Insights and Ideas
- Certified Robustness: The most significant "aha" moment is that you don't have to trade off performance for robustness. By using RENs, the stability guarantees hold throughout the entire training process, regardless of whether the optimizer has fully converged.
- Filtering is Part of the Model: Instead of treating pre-filtering as a separate preprocessing step, BoBDeLaN treats the filter and the recurrent network as a single operator. This allows for the derivation of frequency-dependent gain bounds that are much more useful for robotics than a single global gain.
- Smoothness as a First-Class Citizen: In robotics, "smoothness" (low jitter) is often as important as "accuracy." BoBDeLaN demonstrates that by bounding high-frequency sensitivity, you naturally get smoother control without needing to heavily tune cost functions in the MPC.
Key Results
The experimental results highlight a clear advantage over LSTM-based adaptation:
- Noise Rejection: In frequency sensitivity tests (Figure 2), the REN encoder (orange) strictly followed the theoretical low-pass bound, whereas the LSTM (blue) showed a dangerous sensitivity peak around 5 Hz.
- Smoother Motion: In simulations with unseen payloads, BoBDeLaN reduced position errors to 28% of a nominal MPC (compared to 45% for the LSTM). Crucially, it was much smoother, with an acceleration RMS only 1.8x the nominal baseline, while the LSTM was 2.7x.
- Hardware Resilience: During real-robot pick-and-place tasks with injected noise, the LSTM's performance collapsed, with velocity and acceleration errors spiking by over 600% and 1600% respectively. BoBDeLaN remained stable, maintaining its tracking performance with only minor degradation.
Limitations and Future Work
- Filter Complexity: The current implementation uses a simple low-pass filter. The authors suggest that future work could use more complex filters (like notch filters) to target specific mechanical resonances or structural vibrations.
- Lipschitz Assumptions: The theoretical bounds on torque prediction rely on a Lipschitz continuity assumption for the dynamics model. While this is empirically validated, enforcing this Lipschitz constant during training would provide even stronger "hard" guarantees.
- Broader Applications: The authors aim to extend this framework to other physics-encoded models, such as those for quadrotors (handling aerodynamic disturbances) or legged robots (handling contact-related noise).
Summary updated