Back To BlogsBlog by Aashish Kanted – Aashish has been working on sensor integration and control algorithms for autonomous systems at Ati motors. He is a dual major in Physics and Electrical Engineering from BITS, Goa. His previous projects involved developing a robotic arm and a thrust vector controller for rockets.
Human experiences that are the perfect blueprints for better Robots:
Ever watched a seasoned factory operator drive the same route hundreds of times? They don’t just follow the path — they learn it. Each trip refines their sense of optimal speed, cornering smoothness, and timing, while also preparing them for the unexpected.
At Ati Motors, we apply this very principle to improve motion tracking in our Sherpa line of AMRs through an approach called Learning Model Predictive Control (Learning MPC or LeMPC) — a method that enables robots to refine their control strategy over time for improved precision, safety, and throughput.
From Prediction to Adaptation with LeMPC
In modern control systems, executing a planned trajectory is necessary, but adaptive refinement through learning from operational data and experience is essential for achieving robust and optimal performance.
MPC (Model Predictive Control)
Great for Planning, Not for Learning
- Follows trajectory
- Plans optimal actions
- Adjusts control command
- Never learns from past solutions
LeMPC (Learning MPC)
Robots That Get Smarter with Every Lap
- Follows trajectory
- Plans optimal actions based on previous model(s)
- Adjusts control command & model parameters
- Learns from past solutions
- Refines model and control strategy
Challenges of MPC Addressed by LeMPC
Traditional MPC assumes that the robot motion model accurately represents reality. However, real-world factory and environment conditions often introduce discrepancies that fixed models fail to capture, such as
- Variable floor friction across factory zones.
- Mechanical wear leading to wheel slip or actuator lag.
- Changing payload mass and distribution.
- Temperature and humidity affecting grip and surface adhesion.
These unmodeled factors cause model mismatch, degrading tracking accuracy and increasing takt times. Over time, this leads to performance drift — a few centimeters of accumulated error that reduces overall efficiency.
Learning MPC addresses these limitations through adaptation. Instead of relying solely on a pre-defined physics model, it continuously learns from real-world experience to improve its internal understanding of the system dynamics.
The Learning Process
Simulation Insights
Our simulations compared traditional MPC and Learning MPC under challenging, heavy-payload conditions.
Tracking Accuracy: Both started with ~20 cm error. LeMPC improved to ~5 cm by lap four—a 4x improvement.
(To view comparison of tug tracking the green dotted line using MPC and Learning MPC)
Speed and Efficiency: LeMPC averaged 0.7 m/s vs. MPC’s 0.6 m/s, finishing laps 20 seconds faster.
(Speed comparison using MPC and Learning MPC)
Why LeMPC Matters
- Built on Physics, Enhanced by Data: Learning MPC does not discard physical modeling; it builds upon it. The system starts with a well-established dynamic model and incrementally refines it through empirical learning — achieving both robustness and data efficiency.
- Adaptive to Operating Conditions: Factories are dynamic. Surfaces wear, payloads vary, and layouts evolve. Learning MPC enables AMRs to self-calibrate to their environment, continuously aligning their internal model to the actual operating context.
- Environment-Specific Optimization: Each facility has its own “personality” — floor types, traffic density, and corner geometries. Learning MPC empowers robots to optimize their behavior per environment, rather than relying on generic tuning that may underperform locally.
The Potential of LeMPC
Beyond trajectory tracking, Learning MPC opens up fascinating new avenues:
- Predictive Maintenance: Recurrent patterns in learned model corrections may serve as early indicators of mechanical wear, motor degradation, or sensor drift — allowing for proactive intervention.
- Cross-Robot Knowledge Transfer: Insights from one robot’s experience may be transferable to others operating under similar conditions, accelerating learning fleet-wide [8].
- Hybrid Control Strategies: Combining physics-driven MPC with reinforcement or imitation learning techniques can further enhance robustness and adaptability.
Conclusion
Learning MPC represents a shift in how we think about robot control. Instead of engineering systems that merely follow predefined trajectories, we are building systems that learn to follow better with every run.
For Ati’s AMRs, this means continual self-improvement — mastering optimal speeds, handling varying payloads, and adapting to evolving environments. In essence, each mission becomes a lesson, and every lap, an opportunity for refinement.
The future of factory automation isn’t just about robots that work — it’s about robots that learn to work better, faster, and safer over time.
References
- Rosolia, U. & Borrelli, F. (2019). “Learning How to Autonomously Race a Car: a Predictive Control Approach.” arXiv preprint arXiv:1901.08184v412.
- Xue, H., Zhu, E. L., Dolan, J. M., & Borrelli, F. (2024). “Learning Model Predictive Control with Error Dynamics Regression for Autonomous Racing.” arXiv preprint arXiv:2309.10716v2.
- Rosolia, U. & Borrelli, F. (2017). “Learning Model Predictive Control for Iterative Tasks.” IEEE Transactions on Automatic Control, 63(7), 1883–1896.
- Carneiro, T. F. d. S. C. (2021). MPC Motion Control of an Autonomous Car (Master’s thesis). Instituto Superior Técnico, Universidade de Lisboa.
- Rosolia, U. & Borrelli, F. “Learning Model Predictive Control for Iterative Tasks. A Data-Driven Control Framework.”
- Pinho, J., Costa, G., Lima, P. U., & Botto, M. A. “Learning-Based Model Predictive Control for Autonomous Racing.”
- Borrelli, F. (n.d.). “Iterative Learning Model Predictive Control.” Lecture at University of California Berkeley, USA.
