Ray-tracing Probabilistic Trajectory Learning

• RPTL is a framework that combines ray-tracing with probabilistic modeling to infer and reconstruct complex motion trajectories from sparse 2D or geometric inputs.
• It leverages generative models and neural networks to enable efficient robot learning-from-demonstration and high-efficiency sampling in radio channel propagation.
• Key innovations include geometric projections, learned probability flows, and scene-invariant embeddings that dramatically reduce computational complexity in high-dimensional path spaces.

Ray-tracing Probabilistic Trajectory Learning (RPTL) designates a class of algorithms that combine ray-tracing principles with probabilistic modeling to efficiently learn, sample, or reconstruct complex motion or path distributions, particularly where trajectories are governed by sparse or indirect 2D or geometric observations. RPTL frameworks have been applied to problems ranging from robot Learning-from-Demonstration (LfD) via diagrammatically-specified trajectories (Zhi et al., 2023) to high-efficiency sampling of valid propagation paths in electromagnetic ray tracing for radio channel modeling (Eertmans et al., 2 Mar 2026, Eertmans et al., 2024). RPTL's core innovation is to treat the search for valid or intended trajectories as probabilistic inference over high-dimensional path spaces, aided by generative models and (often) neural networks, leveraging ray-tracing or geometric projections to connect observations or constraints across different spatial representations.

1. Foundational Motivation and Context

Traditional trajectory learning and path-finding in robotics and physics-based simulation often rely on either (i) direct physical demonstrations (e.g., kinesthetic guidance of robot arms), (ii) exhaustive enumeration of all candidate paths (e.g., in classical point-to-point ray tracing), or (iii) direct function learning from prior samples. Each of these approaches is constrained by either high cognitive or hardware interface load (as in kinesthetic/teleoperation LfD), or severe computational intractability due to the exponential growth in feasible trajectory/path candidates with scene complexity or interaction order (as in multi-bounce ray tracing) (Zhi et al., 2023, Eertmans et al., 2 Mar 2026, Eertmans et al., 2024).

RPTL replaces such approaches with frameworks where samples of trajectories are inferred probabilistically—either from indirect observations, such as 2D sketches, or by learning distributions that concentrate on valid or intended solution subsets. This permits orders-of-magnitude reduction in computational cost and interface complexity while retaining fidelity to the underlying intent or physics.

2. Methodological Frameworks

Diagrammatic Teaching and Robot LfD

In the setting of robot skill acquisition via LfD, RPTL enables users to sketch desired end-effector trajectories over one or more calibrated 2D images of the workspace (Zhi et al., 2023). These multi-view sketches are processed as follows:

• From Strokes to 2D Densities: Each user's sketch yields time-stamped pixel traces $\{(x_k, y_k, t_k)\}$, which are used to train a conditional normalizing flow $g^j$ mapping $(t, x, y)$ to a smooth density $p^j(u \mid t)$ per view $j$.
• Ray-Tracing to 3D: For each pixel $u$ above a density threshold at time $t$, a camera-calibrated 3D ray $r^j_u(\lambda)$ is defined. The intersection $R_t = R^1_t \cap R^2_t$ of likely regions from multiple views isolates 3D space consistent across all sketches.
• Probabilistic Trajectory Fitting: The union of $(t_i, x_i)$ samples in 3D is used to fit a Probabilistic Movement Primitive (ProMP) model—i.e., a Gaussian-weighted linear basis expansion $g^j$0. The weights $g^j$1 are learned by maximizing likelihood over recovered 3D samples.
• Trajectory Synthesis: New 3D trajectories are sampled from the learned ProMP, incorporating uncertainty and, if needed, constraints such as pre-specified start positions.

Ray Path Sampling in Propagation Modeling

For radio channel propagation, RPTL replaces brute-force path enumeration with a learned probabilistic sampler capable of prioritizing valid paths (Eertmans et al., 2 Mar 2026, Eertmans et al., 2024):

• State/Action MDP Formulation: The path construction is treated as a finite-horizon Markov Decision Process (MDP), with states representing partial geometric paths and actions extending paths via surface/direction selection.
• Generative Flow Networks (GFlowNets): Global nonnegative flows are assigned to each possible extension of a partial path, such that the induced path-distribution samples with probability proportional to reward $g^j$2 (e.g., geometric validity or ray gain). Flows are learned to satisfy flow-conservation constraints, often via squared error or trajectory-level KL losses.
• Neural Network Architecture: Scene geometry is encoded via permutation- and transformation-invariant neural modules (e.g., E(3)-equivariant MLPs, DeepSets pooling), with candidate-object selection at each step produced by per-object MLPs modulated by state features.
• Sampling Efficiency: At inference, the sampler recursively draws path-extensions proportional to learned flow values, focusing sampling on high-reward (i.e., valid) trajectories.

3. Key Algorithmic Components and Innovations

RPTL implementations rely on several technical innovations to address reward sparsity, combinatorial branching, and geometric consistency:

Component	Description	Paper Reference
Experience Replay Buffer	Buffers rare valid trajectories for replay during GFlowNet training	(Eertmans et al., 2 Mar 2026)
Uniform Exploration Policy	Probability mixing with uniform action to avoid overfitting	(Eertmans et al., 2 Mar 2026)
Physics-based Action Masking	Pre-masks infeasible extensions for geometric or physical validity	(Eertmans et al., 2 Mar 2026)
Scene-conditional Embeddings	Neural featurization invariant to object geometry and permutation	(Eertmans et al., 2024)

These components are essential in highly-sparse situations—e.g., multi-bounce ray tracing—where valid solutions are exponentially rare.

4. Quantitative Evaluation and Comparative Performance

Learning-from-Demonstration for Robotics

In simulated (PyBullet) and real-robot experiments, RPTL achieves substantial improvements over baseline methods. In three robotic manipulation scenarios with five demonstrations per view, RPTL yielded mean discrete Fréchet distances (MFD ×10⁻²) of (3.1±0.2, 3.9±0.2, 5.3±0.9) and Wasserstein distances (2.6, 2.9, 3.8) versus baselines of 10.6–33.2 (MFD) and 6.3–19.1 (WD). These results demonstrate more faithful reconstruction of spatial-temporal intent compared to linear interpolation or nearest-neighbor view-reprojection (Zhi et al., 2023).

On real hardware (fixed-base and quadruped-mounted manipulators), RPTL robustly transferred complex multi-segment trajectory skills from sketches, validating the end-to-end “from sketch to skill” pipeline.

Propagation Path Sampling

In 3D urban canyon simulations (N≈300 facets), RPTL delivered an accuracy of ≈40% and a hit-rate of ≈80% for first-bounce (K=1) paths, compared to ≈3% accuracy with random sampling at equal candidate count. For K=2 (double-bounce), accuracy was ≈15%, hit-rate ≈30%, while random baselines stayed at ≈0.03%. These results demonstrate an order of magnitude reduction in evaluations while retaining most valid solutions (Eertmans et al., 2024). With GFlowNet architectural components, coverage ≥0.95 at 10× GPU/1000× CPU speedup over exhaustive search was demonstrated for 4th-order interactions (Eertmans et al., 2 Mar 2026).

5. Theoretical Properties and Limitations

RPTL inherits key strengths and some open challenges from its hybrid geometric-probabilistic approach:

• Computational Complexity: Exponential scaling in the number of paths or trajectory candidates is reduced to linear growth in sample count and per-step cost (O(M·K·N) for M samples, K bounces, N objects), as opposed to O(N^K) for exhaustive enumeration (Eertmans et al., 2024).
• Generalization: Scene-invariant and permutation-invariant encodings enable RPTL to generalize across novel geometries. Conditioning on learned environment embeddings allows meta-training and zero-shot generation in new scenes (Eertmans et al., 2 Mar 2026).
• Reward Sparsity: For high-order interactions, valid-reward trajectories become vanishingly rare, slowing convergence and necessitating architectural interventions such as replay buffers or hierarchical/continuous reward schemes (Eertmans et al., 2024).
• Physical Modeling Scope: Existing RPTL implementations in radio propagation have been limited to specular reflection; extensions to diffraction, diffuse scattering, and polarization remain open (Eertmans et al., 2024).
• Hardware Scalability: Scalability to city-scale environments (N>10⁶) will require further architectural upgrades, e.g., sparse-attention, hierarchical sub-sampling, or multi-resolution approaches (Eertmans et al., 2024).

6. Extensions and Research Directions

Beyond current applications, RPTL frameworks are being extended to:

• Learning full distributions of complex trajectory parameters (delay, angle, gain) for channel-model fitting by incorporating importance weights into reward functions (Eertmans et al., 2 Mar 2026).
• Meta-learning across heterogeneous environments, leveraging GFlowNet conditional policies to allow zero-shot transfer to previously unseen domains (Eertmans et al., 2 Mar 2026).
• Variable-length or infinite-horizon path generation tasks, increasing applicability to diffuse or stochastic processes (Eertmans et al., 2 Mar 2026).
• Enhanced reward and training strategies, including partial reward shaping and hierarchical policies, to further ameliorate convergence under sparse-reward regimes (Eertmans et al., 2024).

This suggests RPTL constitutes a foundational approach at the intersection of geometric modeling and probabilistic generative learning, with ongoing research aimed at expanding its scalability, modeling capacity, and application spectrum.