Real-Time Motion Planning with Dynamic Hazards: Classical vs. Learning-Based Methods

arXiv:2610.12249 · cs.RO, cs.AI · Submitted 2026-10-08 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: I'm Rosa, and with me are Dev and Taro, guest researcher.

Dev: Today's paper: "Real-Time Motion Planning with Dynamic Hazards".

Rosa: The gist: Uncertainty in obstacle evolution, rather than partial observability, is the key factor that changes planning difficulty and determines which paradigm is practically effective in real time.

Dev: First, who's behind it and why it matters.

Title and authors: Rosa: So we're looking at this paper called "Real-Time Motion Planning with Dynamic Hazards: Classical vs. Learning-Based Methods." It’s by Eran Iceland and his team at the Hebrew University of Jerusalem.

Dev: Yeah, it’s a comparison study between classical planning methods and learning-based methods in these dynamic environments. Basically, they build a benchmark where both types of planners face the exact same settings for visibility and obstacle movement.

Taro: It seems like they set up this testbed to see how well each approach handles the challenges when things start changing unpredictably, which is what we care about a lot in autonomy research.

Rosa: Exactly. The core idea here is that uncertainty in how obstacles move is way more important than just not seeing everything at once, which is partial observability. The paper sets up four different scenarios to test this out.

Dev: Right, they're varying visibility—full versus limited—and obstacle dynamics—deterministic versus stochastic. This lets them isolate what causes the planning difficulty in real time and which system works best under those specific conditions.

Taro: I think it’s interesting how they structured those regimes. They start with full visibility and constant angular velocity, then move into limited visibility, and then introduce the stochastic changes to the obstacle speeds.

Rosa: That’s right. It really forces you to ask what kind of uncertainty actually breaks the planning system when you have a tight deadline.

Dev: And they compare this setup against several classical planners, like A* and GBFS, as well as a reinforcement learning planner using PPO for training.

Taro: The way they set up those baselines is important because it gives us a solid ground truth to compare the performance of the AI approach against established algorithms.

Rosa: And what we see immediately is that in the deterministic regime, the classical planners, A* and GBFS, actually outperform the reinforcement learning approaches when visibility is full.

Dev: It's interesting because it shows that when everything is predictable and you can see it all, a traditional search planner can find a better path than what a trained policy might produce.

Title and authors: Taro: That makes sense if the environment is fully known and deterministic; the search algorithm has perfect geometric guarantees in that setting.

Rosa: Now, they move into the stochastic obstacle dynamics regime, and this is where things get really telling about uncertainty versus partial observability.

Dev: When those stochastic changes are introduced, the classical methods start showing a lot of sensitivity to planning time. They can fail almost completely if the time budget is too small.

Taro: That suggests that in uncertain environments, running out of time is a huge factor that causes failure for deterministic planners.

Rosa: But here’s the contrast: the reinforcement learning methods maintain stable performance across those regimes while keeping their paths shorter and using less computational power when the obstacle evolution is stochastic.

Dev: That points to a key trade-off we see often—learning-based systems are robust in uncertainty, even if they aren't as fast or precise as a perfect classical planner in the ideal case.

Taro: So, under uncertainty about obstacle movement, the learning approach seems more practical for real time operation than the classical search methods.

Rosa: That’s what they conclude: uncertainty in obstacle evolution is the dominant factor affecting planning difficulty and determines which paradigm is practically effective in real time. This whole study on "Real-Time Motion Planning with Dynamic Hazards: Classical vs. Learning-Based Methods" really highlights that distinction.

Dev: And they do point out some important things in their ablation study regarding how the RL planner was tuned. They found that changing the reward structure—specifically balancing sparse and dense rewards—had a clear effect on path length under full visibility with fixed sprinkler angular velocity.

Taro: That suggests we can tune the learning agent to prioritize different things depending on what we need from it, like local efficiency versus just getting to the final goal reliably.

Title and authors: Rosa: And they also found that for the classical stochastic planner, its main weakness was actually related to its decision interval settings within the UCT framework. That’s a specific tuning knob that matters for those traditional search algorithms when things are random.

Dev: So we have these different ways to look at this problem: one side is relying on fast, safe online planners under uncertainty, and the other relies on learning systems that adapt quickly to unpredictable changes.

Taro: It makes me wonder how we can combine those two ideas—maybe using a classical planner to set the high-level path and then an RL policy to handle the immediate local adjustments when things go stochastic.

Rosa: That sounds like a very interesting direction for future work, exploring how to leverage the strengths of both methods for better real-time autonomy.

Dev: It’s definitely something worth looking into, especially since we saw how RL methods managed latency better under those stochastic conditions compared to the online UCT baseline in this paper.

Taro: So, to wrap up on this "Real-Time Motion Planning with Dynamic Hazards: Classical vs. Learning-Based Methods," the big picture is that when the environment evolves stochastically, dealing with that evolution uncertainty is the main challenge for any motion planning system we build today.

Rosa: It really shows us that just being able to see everything at once isn't enough; understanding how things are changing over time and space is what separates a successful planner from a failed one in real time applications.

Dev: And this comparison between classical tools and learning planners under those four different conditions gives us a clearer picture of when we should expect high success rates versus low latency.

Taro: I think the implication for autonomy is that the choice of planning paradigm needs to be dictated by the nature of the uncertainty present in the hazard field, not just by whether we have perfect visibility or not.

Rosa: That’s what this paper on "Real-Time Motion Planning with Dynamic Hazards: Classical vs. Learning-Based Methods" really gets at, forcing us to be more nuanced about how we design these systems for the real world.

The paper's summary: Rosa: So, to recap, this paper sets up a testbed where they compare old-school search planners against modern reinforcement learning agents when dealing with moving obstacles and changing visibility in a robot's pathfinding mission.

Dev: Right, it’s not just about whether you can see everything or not; it’s really about how fast that environment is changing, which makes planning hard in real time.

Taro: The main point they drive home is that uncertainty in how the obstacles are evolving—that unpredictability—is what actually breaks the classical methods when things get stochastic, more than just having a bit of partial visibility.

Rosa: Exactly. They show that if you introduce random changes to how fast those hazards move, the deterministic planners start failing way faster than they would have under normal conditions.

Dev: And that’s a huge concern for us on the control side; we see this sensitivity to planning time in our loops, and this paper quantifies exactly why it happens when the dynamics are uncertain.

Taro: The results show that while classical planners like A* and GBFS can give you a perfect path in the fully known, deterministic case, they become extremely brittle when those obstacle speeds start fluctuating randomly.

Rosa: But what’s interesting is that the reinforcement learning approach doesn't crash; it stays stable across all four of these different conditions—full visibility, limited visibility, and both dynamic and static obstacle movement.

Dev: That stability is key for us because it means we can actually use the AI planner under those stochastic dynamics without worrying about catastrophic failure just because the environment got a little weird.

Taro: They also found that the RL agent can keep its paths shorter than what the classical planners produce when things are moving randomly, which suggests a better trade-off for real-world execution.

Rosa: So, it comes down to this: when you’re in a situation where the obstacle evolution is unpredictable, you have to choose between a fast but brittle search algorithm or an AI approach that handles that uncertainty more gracefully.

Dev: And what this means for us on the engineering side is that we need to design our systems to dynamically switch between those two paradigms based on how much uncertainty we detect in the input stream.

Taro: It opens up a whole new area of research about how we can build planners that are inherently robust to dynamic evolution, not just reactive to being partially blocked.

The paper's improvements: Tom: So, we're looking at how they suggest fixing these issues in their study about motion planning under dynamic hazards.

Rosa: Basically, they show that you can actually make the system smarter by dynamically picking which method to use based on what’s happening in the environment.

Dev: That’s a big deal for us because it means we don't have to commit to just one planning approach when things get messy; we can switch modes.

Taro: It suggests that if the uncertainty is really about how the obstacles are evolving, using a learning-based method might be way more effective than sticking with a deterministic search planner.

Rosa: Right, they argue that having an adaptive system lets you handle those "regime shifts" where things suddenly get unpredictable without losing success rates.

Dev: And on the latency side, they found that if you use those trained reinforcement learning policies under stochastic dynamics, the per-decision delay stays really low even when the environment is changing fast.

Taro: That addresses a major issue we see with learning models; we get these long deliberation times, but this paper shows that tuning the reward structure can keep things focused and short.

Rosa: They demonstrated that by balancing sparse rewards for finishing the mission against dense rewards for local movement, you can guide the AI to be efficient without sacrificing the final outcome reliability.

Dev: So, it’s not just about one planner being better; it’s about building a flexible system that knows when to trust the search algorithm and when to let the learning policy take over.

Taro: The implication for autonomy is that we need planning frameworks that can sense the nature of the uncertainty and adjust their strategy accordingly, instead of relying on a fixed setup.

Rosa: And they also pointed out some limitations, which is important; they noted that while this works well in simulation or controlled scenarios, moving it to a completely unstructured real-world field is still a big hurdle.

Dev: Yeah, the hardware environment they used—that specific Azure machine—is part of the caveat because real-world performance will depend on how robust the AI handles sensor noise and unpredictable physical interactions.

Taro: So what we’re left with is a framework that uses learning to handle the chaos while keeping classical methods ready for those moments when everything is predictable, which sounds like a solid path forward.

Conclusion: Rosa: So, we're wrapping up this look at "Real-Time Motion Planning with Dynamic Hazards: Classical vs. Learning-Based Methods." It really boils down to this—uncertainty in how obstacles move, not just not seeing everything at once, is what dictates which planning method actually works in real time.

Dev: Right, it’s a comparison between the speed and precision of classical search algorithms and the stability of reinforcement learning agents under changing conditions.

Taro: The big picture here is that for any autonomous system facing dynamic hazards, you can't just pick one planning tool; you need something that adapts to the level of unpredictability.

Rosa: Exactly. They showed that when the environment gets random, like those stochastic speed changes, the RL approach maintains a much better balance between path quality and low computational cost compared to traditional methods.

Dev: From an engineering standpoint, this means we can design systems that dynamically switch their planning strategy based on whether they sense high uncertainty or if the environment is relatively predictable.

Taro: It also gives us a clear direction for future work: figuring out how to integrate those two worlds, so we get the geometric guarantees of A* in good situations and the adaptive power of AI when things go sideways.

Rosa: And they did point out that while the simulation results are solid, testing this kind of adaptability in a messy real world is still the next big step for field robotics.

Dev: Yeah, and we need to keep an eye on how well those policies handle unexpected sensor noise when they're running at high loop rates.

Taro: I think the paper gives us a strong framework to test those kinds of robust behaviors in simulations before we even deploy hardware.

Eran Iceland, *Alexander Tuisov, *Oren Gal, *Ariel Barel†, *Alfred M. Bruckstein

School of Engineering and Computer Science, The Hebrew University of Jerusalem · Faculty of Data and Decision Science, Technion Israeli Institute of Technology · Hatter Department of Marine Technologies, University of Haifa

cs.RO, cs.AI

Submitted: 2026-10-08

Updated: 2026-10-08

The gist: The gist: Uncertainty in obstacle evolution, rather than partial observability, is the key factor that changes planning difficulty and determines which paradigm is practically effective in real time.

Key concepts

Uncertainty in Obstacle Evolution
This refers to the unpredictability of how moving hazards change their positions or movement patterns over time. The paper argues that this uncertainty, rather than just not seeing everything (limited visibility), is the primary driver making motion planning hard in dynamic environments.
Full Visibility vs. Limited Visibility
This describes whether the robot can see all relevant parts of its environment at once (full visibility) or only a partial view. The study tested both scenarios to see how this affects planning, especially when combined with different obstacle dynamics.
Stochastic Obstacle Dynamics
This means the obstacles move in a random or probabilistic way rather than following a fixed, predictable path. This tests the system's ability to handle true randomness in obstacle movement, which is crucial for real-world applications.

Terminology

Summary

The gist: Uncertainty in obstacle evolution, rather than partial observability, is the key factor that changes planning difficulty and determines which paradigm is practically effective in real time.

Problem Formulation and Regimes

The study constructs a unified testbed for motion planning in dynamic environments by varying two axes: visibility (full vs. limited) and obstacle dynamics (deterministic vs. stochastic) The problem is formalized under full visibility as a deterministic formulation via an XY T lifting This setting involves the robot advancing monotonically along T at unit rate while time advances monotonically, creating a tight coupling between spatial motion and temporal progression The four regimes considered are (i) full visibility with constant angular velocity obstacles, (ii) limited visibility with constant angular velocity obstacles, (iii) full visibility with stochastically changing angular velocities, and (iv) limited visibility with stochastic dynamics.

Classical Planning Method Comparison

Classical baselines are implemented tailored to the regime’s assumptions. In the deterministic obstacle angular velocity regime, A and GBFS are used as high-quality and fast satisficing baselines, respectively. Under limited visibility, a D-like replanning framework is employed where A and GBFS serve as online replanning components rather than globally optimal planners. For the stochastic obstacle angular velocity regime, Upper Confidence bounds applied to Trees (UCT) are used as a stochastic planning baseline.

Reinforcement Learning Planner

The RL planner is modeled as an episodic agent executed in a closed loop. Policies are trained using PPO, chosen for its stability under sparse and delayed feedback in dynamic environments and its common usage in image-based control with discrete action spaces. The evaluation involves running policies on scenarios sampled from the same distribution as training, executed without further learning or adaptation.

Experimental Results and Key Findings

In the deterministic regime, classical planners clearly outperform RL-based approaches, particularly under full visibility. Both A and GBFS achieve perfect success rates, with A producing near-optimal trajectories and GBFS trading optimality for speed. When stochastic obstacle dynamics are introduced, classical methods exhibit strong sensitivity to the available planning time: small timeouts lead to near-complete failure, while larger budgets improve success rates but result in significantly longer and more conservative trajectories. In contrast, RL methods achieve stable performance across regimes while maintaining substantially shorter paths and low computational cost under stochastic obstacle evolution.

Conclusion on Dominant Factor

The results demonstrate that the degradation of classical methods in the stochastic regime is driven by uncertainty rather than limited visibility, as performance drops sharply even under full visibility once stochastic dynamics are introduced. This comparison shows that uncertainty in obstacle evolution is the dominant factor affecting planning difficulty and determines which paradigm is practically effective in real time. The RL policy consistently outperforms the online UCT baseline under stochastic obstacle evolution in latency, success rate, and path quality. This indicates that uncertainty in obstacle evolution, rather than partial observability, is the key factor that changes planning difficulty and determines which paradigm is practically effective in real time >.

Ablation Study Insights

The ablation study showed that for the classical stochastic planner, the main sensitivity was the UCT decision interval. For RL methods, reducing the completion reward from 25 to 5 while increasing the progress reward had a clear effect, shortening paths under full visibility with fixed sprinkler angular velocity but reducing success under full visibility with deterministic angular velocity. This suggests that stronger dense rewards encourage locally efficient motion, whereas a larger terminal reward is important for reliable task completion. The average wall-clock training throughput was approximately 550 FPS, corresponding to about 2 million environment steps per hour. All experiments were conducted on an Azure virtual machine equipped with 8 vCPUs (AMD EPYC 7V12 processor, 2.45 GHz) and 54 GB RAM, under a Microsoft Azure hypervisor. The classical planning algorithms (A∗, GBFS, and UCT) were executed on the CPU only, each restricted to a single CPU core. The RL approach was evaluated under two configurations: (1) CPU-only execution, using a single CPU core, and (2) GPU-accelerated execution, using a single NVIDIA Tesla T4 GPU. All experiments were conducted on an otherwise idle virtual machine to minimize variability. The classical baselines are representative methods for controlled comparison with the RL planner. The test environment consists of planar domains populated with rotating sprinkler-like hazards that generate time-varying forbidden regions via sweeping angular sectors. The agent is considered to have reached the goal if ∥(xt, yt) − g∥ < 1. The robot is modeled as a holonomic point moving at fixed speed and bounded turn rate in the plane while time advances monotonically, i.e. the agent must advance monotonically along T at unit rate. The state-space search problem is defined by an initial state, a successor function, a goal test, and a path-cost function. The agent observes two ego-centered RGB images of size 3 × 101 × 101 at scales 0.5 and 2. The multi-scale representation enables both local precision and global context. Sprinklers are angular sectors; non-flight zones (NFZs) are defined as a uniform inflation of each sector with margin 1. The agent observes only part of the spacetime graph, which can be discovered during execution. A discrete action space of 7 steering actions is defined by at = ∆θt ∈ (−45◦, −30◦, −15◦, 0◦, 15◦, 30◦, 45◦). The agent moves one unit per timestep in the direction of its heading. The agent is considered to have reached the goal if∥(xt+1, yt+1) − (xt, yt)∥ = 1. The visualization of the initial state can be seen in Figure 1. The visualization of the corresponding trajectories can be seen in Figure 2. Table 1 reports A∗ offline planning time as mean and standard deviation over 1,000 deterministic full-visibility scenarios. Table 3 shows the success rate versus maximal replan time under limited visibility for (a) deterministic and (b) stochastic dynamics. Table 4 reports A∗ offline planning time as mean and standard deviation over 1,000 deterministic full-visibility scenarios. The distribution is strongly right-skewed: most runs finish within tens of seconds, while rare hard instances produce a long tail, inflating the standard deviation reported in Table 1. The results support a single dominant takeaway: uncertainty in obstacle dynamics, rather than partial observability, is the key factor that changes planning difficulty and determines which paradigm is practically effective in real time >.

--- Page 1 ---

Real-Time Motion Planning with Dynamic Hazards: Classical vs. Learning-Based Methods Eran Iceland Alexander Tuisov Oren Gal Ariel Barel Alfred M Bruckstein These authors contributed equally to this work.

Improvements for AI systems

  1. The improved system can robustly handle real-time motion planning in dynamic hazard fields by dynamically selecting between classical and learning-based methods based on environmental uncertainty. This allows for a clear regime shift in performance, ensuring high success rates when uncertainty is dominated by obstacle evolution, as the paper states: uncertainty in obstacle evolution, more than partial observability, is the dominant factor determining which planning paradigm is practically effective for the problem at hand.

  2. The system can operate with significantly reduced per-decision latency by leveraging trained Reinforcement Learning policies under stochastic dynamics. This directly addresses the trade-off observed: RL methods achieve stable performance across regimes while preserving very low per-step latency, yielding a favorable success-quality-latency tradeoff under uncertainty.

  3. The system can maintain near-perfect path quality and success rates in deterministic environments by employing classical A∗ or GBFS planners. This is achieved because In the deterministic regime, classical planners achieve near-perfect success and better path quality, allowing for precise geometric guarantees when the environment is fully known.

  4. The system can execute online replanning strategies effectively under limited visibility by using D°-like frameworks. This capability allows the agent to address partial observability by using methods like A∗ and GBFS are online replanning components rather than globally optimal planners, which maintains high performance even when information is revealed during execution.

  5. The system can be optimized for specific task objectives by tuning the reward structure in a Reinforcement Learning setup. By adjusting rewards, the system can be guided to prioritize local efficiency or overall mission completion, as demonstrated by how changing the balance between sparse and dense rewards had a clear effect on path length and success rates.

Sources

Related papers