Obstacle Avoidance Under Uncertainty — Safe MPC with Imperfect Perception

Public reference use case (by others): This page summarizes publicly available reference implementations and papers. Not client results of Dr. Rafal Noga.

Why This Matters (Executive Summary)

• Real-world mobile systems — warehouse AGVs, collaborative robots, drones, outdoor vehicles — operate with imperfect maps, noisy sensors, and unpredictable moving obstacles. Ignoring this uncertainty leads to either collisions or excessive conservatism (constant emergency stops).
• MPC formulations with chance constraints or robust margins provide a principled way to balance throughput against safety: the controller explicitly reasons about what it does not know, rather than assuming perfect perception.
• Published results include 57% faster mission completion for drone swarms vs. reactive baselines, smooth obstacle avoidance at 11.5 km/h on GPU-accelerated platforms, and scenario-based NMPC outperforming fixed-path safety strategies in human-robot collaboration.
• For DACH operations with mobile platforms in shared or cluttered environments, this pattern reduces unplanned stops while maintaining quantifiable safety guarantees.

The Design Pattern Explained

Classical obstacle avoidance assumes the controller knows exactly where obstacles are and exactly where the robot is. In practice, neither is true: sensors have noise and latency, maps are incomplete, and humans or other vehicles move unpredictably. This pattern addresses the gap by treating uncertainty as a first-class design input in the MPC formulation.

Two main strategies exist:

1. Deterministic margins (robust MPC): Obstacles are inflated by a safety buffer derived from worst-case uncertainty bounds. Simple to implement, but conservative — the robot behaves as if the worst case is always true.

2. Probabilistic chance constraints: The collision probability is bounded to an acceptable level (e.g., < 0.1%), yielding tighter but statistically justified clearances. This requires a probabilistic model of the uncertainty (e.g., Gaussian position error) but allows significantly higher throughput.

A key insight from several reference implementations is that control actions affect perception quality: faster motion degrades sensor accuracy (motion blur, reduced integration time), creating a feedback loop between planning and sensing. Advanced formulations account for this coupling explicitly.

The architecture typically involves: perception (lidar, cameras, depth sensors) producing obstacle estimates with uncertainty, prediction (motion models for dynamic obstacles), MPC optimization (with inflated or chance constraints), and safety monitoring (independent watchdog layer).

Applications & Reference Implementations

UAV NMPC with Dynamic Obstacles — Aerial Navigation

An IEEE RA-L publication demonstrated nonlinear MPC for UAV navigation with dynamic obstacles, using a 50 ms sampling time and a 2-second prediction horizon for both obstacle trajectories and the NMPC optimization. The solver (PANOC via OpEn) used a penalty method for constraint handling, enabling real-time solutions in laboratory experiments with moving obstacles. The 2-second look-ahead gives the UAV enough time to plan avoidance maneuvers while maintaining stable flight, demonstrating that even with modest compute budgets, predictive obstacle avoidance outperforms reactive approaches. ¹

Predictive Drone Swarm Navigation Through Clutter — Multi-Robot Coordination

Research published in Nature Machine Intelligence and covered by WIRED demonstrated predictive control enabling a drone swarm (Crazyflie quadrotors) to traverse a dense “fake forest” of obstacles. The predictive approach completed the mission 57% faster than reactive control baselines (34.1 s vs. 21.5 s in one scenario). The distributed predictive controller reduced both computation time and travel time compared to classical centralized multi-agent MPC, making it scalable to larger swarms. Physical demonstrations validated that the approach transfers from simulation to real hardware. ²

GPU Randomized MPC for Dynamic Obstacle Avoidance — Autonomous Platforms

Muraleedharan et al. implemented randomized MPC on a 1/10-scale RC car, comparing CPU and GPU implementations for dynamic obstacle avoidance. The CPU version achieved 30 Hz and failed (collision) at 5.1 km/h. The GPU version achieved 200 Hz with 1000 trajectory samples, enabling smooth avoidance at up to 11.5 km/h with a 30-step prediction horizon covering approximately 3 m of look-ahead. The massive parallelism of GPU sampling transforms obstacle avoidance from a conservative, low-speed capability into a robust high-speed one. ³

Safe Human-Robot Workspace Sharing via Scenario NMPC — Collaborative Robotics

A Control Engineering Practice paper proposed scenario-based NMPC with probabilistic human-motion predictions for safe human-robot coexistence. Using higher-order Markov chains to generate a scenario tree of likely human motions, the controller modulates robot speed so it can always stop before a collision. Experimental tests on a Kinova Gen3 robot interacting with a human operator showed superior performance compared to both NMPC without human predictions and a fixed-path speed-and-separation monitoring (SSM) strategy. This approach reduces unnecessary slowdowns while maintaining safety guarantees. ⁴

Ballbot Path Following with Obstacle Avoidance — Dynamically Unstable Platforms

A path-following MPC formulation was validated on a ballbot — an inherently unstable robot balancing on a single ball — with obstacle constraints. The MPC ran at 10 Hz on an external computer, handling circular path tracking (1 m radius, 0.25 m/s desired speed) while avoiding obstacles in the workspace. The ballbot case is especially challenging because the robot cannot simply stop: it must continuously balance, so obstacle avoidance and stability must be solved jointly within the same optimization. ⁵

What This Means for Your Operations

For DACH operations running mobile platforms in environments with imperfect perception or moving obstacles, the transferable insights are:

• Start with deterministic margins: inflate obstacles by a conservative buffer derived from your sensor specifications. This is the simplest first step and already outperforms reactive emergency-stop logic.
• Graduate to probabilistic constraints when you have data to characterize your localization and perception uncertainty (e.g., from commissioning runs). This tightens clearances and increases throughput.
• Dynamic obstacle prediction is most valuable in human-robot collaboration and multi-AGV environments. Even simple constant-velocity predictions provide significant benefit over reactive control.
• Common readiness indicators: the platform has encoders and at least one range sensor (lidar, depth camera); a kinematic/dynamic model exists or can be identified; the environment has some structure (warehouse layout, production floor map).

How We Deliver This (Engagement Model)

• Phase 0: NDA + data request — we review your platform sensors, environment maps, and safety requirements (e.g., ISO 3691-4 for AGVs, ISO/TS 15066 for cobots).
• Phase 1: Fixed-scope discovery — uncertainty characterization, constraint formulation (deterministic vs. probabilistic), and solver feasibility assessment on your target hardware.
• Phase 2: Implementation + validation — perception-control integration, MPC solver deployment, simulation testing with injected uncertainty, and on-hardware commissioning.
• Phase 3: Monitoring + training + scaling — safety KPI dashboards, operator training on constraint tuning, and rollout to additional vehicles or cells.

Typical KPIs to Track

• Collision rate (must be zero or below the specified chance-constraint threshold)
• Unplanned stop frequency — reduced stops indicate better predictive planning
• Average clearance margin — tighter margins (while maintaining safety) indicate better uncertainty handling
• Mission completion time — throughput impact of the obstacle avoidance strategy
• Perception-control latency — end-to-end delay from sensor input to actuator output

Risks & Prerequisites

• Uncertainty characterization is essential: the quality of the safety guarantees depends directly on how well you model your sensor and localization uncertainty. Overly optimistic uncertainty models lead to unsafe behavior.
• Compute budget increases compared to obstacle-free MPC: chance constraints or scenario trees add variables and constraints to the optimization. Profile early on target hardware.
• Sensor coverage gaps (blind spots, occlusions) cannot be solved by the controller alone — they require physical sensor placement design.
• Regulatory context: safety-critical applications may require additional certification beyond the MPC design (e.g., independent safety-rated monitoring layers per ISO 13849).

FAQ

Q: How is this different from simply adding a larger safety buffer around obstacles? A: A fixed buffer is worst-case conservative everywhere. Probabilistic chance constraints adapt the effective buffer based on actual uncertainty: tight clearances where perception is confident, wider margins where it is not. This typically yields 20-40% higher throughput.

Q: Can this work with only lidar, or do I need cameras too? A: Lidar alone is sufficient for static and slow-moving obstacles. For fast dynamic obstacles or humans, adding cameras or depth sensors improves prediction quality. The MPC formulation is sensor-agnostic — it consumes obstacle positions and uncertainty estimates regardless of the source.

Q: What about environments that change frequently (e.g., warehouse reconfigurations)? A: The MPC does not depend on a static map — it uses the real-time obstacle estimates from perception. Map changes are handled transparently as long as the perception pipeline detects them. Pre-mapped environments improve performance but are not strictly required.

Q: Is this applicable to slow-moving AGVs or only to fast drones? A: Both. For slow AGVs (< 2 m/s), the compute budget is generous and even simple deterministic margins provide major benefits over reactive e-stops. For fast platforms, the probabilistic formulation becomes essential to avoid excessive conservatism.

Book a 30-Minute Discovery Call

Ready to explore whether uncertainty-aware obstacle avoidance fits your mobile platform?

Dr. Rafal Noga — Independent APC/MPC Consultant

Email me | noga.es

Fixed-scope discovery — NDA-first — DACH on-site available

Public References

Footnotes

1. “Real-time NMPC for UAV navigation with dynamic obstacles” (IEEE RA-L, 2021). https://www.diva-portal.org/smash/get/diva2:1457693/FULLTEXT01.pdf ↩

2. Soria et al., “Distributed Predictive Drone Swarms in Cluttered Environments” (Nature Machine Intelligence, 2021). https://aerial-core.com/pubs/soria_nmi21.pdf ↩

3. Muraleedharan et al., “Randomized MPC for Dynamic Obstacle Avoidance and Autonomous Racing” (IEEE T-IV, 2022). https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=9911860 ↩

4. Cimini et al., “Safe human-robot workspace sharing via scenario-based NMPC” (Control Engineering Practice). https://cse.lab.imtlucca.it/~bemporad/publications/papers/cep-smpc-robot.pdf ↩

5. “Path-following MPC for ballbots with obstacle avoidance” (arXiv, 2020). https://arxiv.org/pdf/2008.10383.pdf ↩