HyPRAP: Risk-Aware Hybrid Prediction Planning

Updated 6 July 2026

HyPRAP is a hybrid planning framework that selectively allocates prediction effort using a Collision Risk Index to route obstacles based on risk.
It integrates hybrid conformal prediction with model predictive control to attach uncertainty bounds, enabling informed collision avoidance.
Empirical evaluations show HyPRAP nearly matches high-accuracy predictors in success rate while reducing computation compared to uniform prediction models.

Hybrid Prediction-based Risk-Aware Planning (HyPRAP) denotes a prediction-based risk-aware path-planning framework for real-time navigation in dense, uncertain environments. In its explicit formulation, HyPRAP uses a hybrid combination of models to predict local obstacle movement, a Prediction-based Collision Risk Index (P-CRI) to evaluate obstacle relevance, hybrid conformal prediction to attach uncertainty bounds to predictor outputs, and model predictive control (MPC) to plan around only the selected high-relevance obstacles (Yang et al., 16 Jul 2025). The term appears explicitly in "Hybrid Conformal Prediction-based Risk-Aware Model Predictive Planning in Dense, Uncertain Environments" (Yang et al., 16 Jul 2025), while several earlier and contemporaneous systems instantiate closely related hybrid prediction–planning patterns without using the same acronym (Dash et al., 23 Feb 2026).

1. Problem setting and architectural motivation

HyPRAP is posed for a single agent evolving under discrete-time dynamics

$\xvect_{t+1} = f(\xvect_t,\uvect_t), \qquad t \in \mathbb{Z}_{\ge 0},$

with state $\xvect_t \in \mathbb{R}^n$ and control $\uvect_t \in \mathbb{R}^m$. The environment contains moving obstacles

$\mathcal{N} \triangleq \{1,\dots,N\},$

but at time $t$ only a sensed subset $\mathcal{N}_t \subseteq \mathcal{N}$ is available, with $N_t = |\mathcal{N}_t|$ (Yang et al., 16 Jul 2025). The formulation assumes a single agent, dynamic obstacles with unknown dynamics, local sensing only, no sensor noise in the formulation, a known static map, and predictor independence in the hybrid conformal section.

The motivating difficulty is computational asymmetry. In dense scenes, predicting every obstacle with a uniformly accurate model is expensive, while predicting every obstacle with a uniformly cheap model enlarges uncertainty sets, shrinks the MPC feasible region, and can make planning overly conservative or infeasible. HyPRAP addresses this by exploiting the fact that the agent does not interact equally with all obstacles. It therefore routes obstacles to different predictors according to predicted planning relevance rather than treating all sensed obstacles uniformly (Yang et al., 16 Jul 2025).

This yields a characteristic decomposition. At each planning cycle, the system observes local obstacles, computes a risk score for each obstacle, routes each obstacle to a predictor according to that score, generates predictions only for the selected subset, attaches conformal uncertainty bounds appropriate to the chosen predictor, and solves MPC with collision constraints only for that selected subset. This suggests a planner in which computational effort is itself allocated by prediction-derived risk, rather than only by geometric proximity.

2. Hybrid composition and selective routing

The defining “hybrid” property of HyPRAP has two layers. First, it uses multiple prediction models with different speed–accuracy tradeoffs. Second, it uses hybrid conformal prediction, meaning that the uncertainty set attached to an obstacle depends on which predictor was selected for that obstacle (Yang et al., 16 Jul 2025).

In the reported simulations, the two predictors are $P^1=\text{LSTM}$ , characterized as higher accuracy and lower computational efficiency, and $P^2=\text{GPR}$ , characterized as higher efficiency and lower accuracy. Obstacles not assigned to either predictor are handled by a constant velocity model so that P-CRI remains defined at the next step, but these low-risk obstacles are not incorporated as full MPC constraints (Yang et al., 16 Jul 2025). The resulting routing logic is risk-tiered: high-risk obstacles use the more accurate model, moderate-risk obstacles use the cheaper model, and very low-risk obstacles are excluded from expensive prediction and planning.

The routing variable is written as $\phi(t,k)$ for obstacle $\xvect_t \in \mathbb{R}^n$0 at time $\xvect_t \in \mathbb{R}^n$1. If $\xvect_t \in \mathbb{R}^n$2, the obstacle is deemed insignificant for current planning. Otherwise it is predicted by $\xvect_t \in \mathbb{R}^n$3, and the selected set is

$\xvect_t \in \mathbb{R}^n$4

with $\xvect_t \in \mathbb{R}^n$5 (Yang et al., 16 Jul 2025). The paper’s online algorithm updates $\xvect_t \in \mathbb{R}^n$6 using the previously planned agent trajectory and the previous obstacle prediction, then forms $\xvect_t \in \mathbb{R}^n$7, runs the chosen predictors for obstacles in $\xvect_t \in \mathbb{R}^n$8, solves MPC, applies the first control, and repeats.

This is not merely a runtime optimization. It changes the planning problem by deciding which obstacles receive accurate forecasts, which receive cheap forecasts, and which are omitted from the active collision-avoidance set. A plausible implication is that HyPRAP should be viewed less as one specific planner than as a modular allocation architecture for prediction effort inside MPC.

3. Prediction-based Collision Risk Index and hybrid conformal prediction

The central routing mechanism is the Prediction-based Collision Risk Index (P-CRI), which scores each obstacle using predicted agent–obstacle geometry over the horizon rather than current distance alone (Yang et al., 16 Jul 2025). For each obstacle $\xvect_t \in \mathbb{R}^n$9, the paper defines the prediction-based approach distance and prediction-based approach time vectors:

$\uvect_t \in \mathbb{R}^m$0

$\uvect_t \in \mathbb{R}^m$1

Here $\uvect_t \in \mathbb{R}^m$2 and $\uvect_t \in \mathbb{R}^m$3 are predicted positions, and $\uvect_t \in \mathbb{R}^m$4 and $\uvect_t \in \mathbb{R}^m$5 are predicted velocities obtained by backward difference from positions.

The scalar index is then

$\uvect_t \in \mathbb{R}^m$6

normalized to $\uvect_t \in \mathbb{R}^m$7. The paper does not specify closed-form choices for $\uvect_t \in \mathbb{R}^m$8 and $\uvect_t \in \mathbb{R}^m$9, but it requires $\mathcal{N} \triangleq \{1,\dots,N\},$ 0 to decrease monotonically, and $\mathcal{N} \triangleq \{1,\dots,N\},$ 1 to increase monotonically when $\mathcal{N} \triangleq \{1,\dots,N\},$ 2, decrease monotonically when $\mathcal{N} \triangleq \{1,\dots,N\},$ 3, and remain continuous at $\mathcal{N} \triangleq \{1,\dots,N\},$ 4; a double-sided exponential is given as an example (Yang et al., 16 Jul 2025). P-CRI therefore combines proximity and approach tendency, rather than using distance alone.

Routing is threshold-based. Given thresholds $\mathcal{N} \triangleq \{1,\dots,N\},$ 5 with $\mathcal{N} \triangleq \{1,\dots,N\},$ 6 and $\mathcal{N} \triangleq \{1,\dots,N\},$ 7,

$\mathcal{N} \triangleq \{1,\dots,N\},$ 8

Higher $\mathcal{N} \triangleq \{1,\dots,N\},$ 9 leads to a more accurate predictor; lower $t$ 0 leads to a cheaper predictor or exclusion (Yang et al., 16 Jul 2025).

Uncertainty quantification is then handled by hybrid conformal prediction. For obstacle $t$ 1 using predictor $t$ 2, the framework seeks

$t$ 3

for each $t$ 4, and extends this to full-horizon coverage with $t$ 5 (Yang et al., 16 Jul 2025). The novelty is not a new conformal score but a predictor-indexed uncertainty calibration: each obstacle uses the conformal radius corresponding to the predictor that routed it.

The paper further notes that confidence should depend on both horizon and the number of predicted obstacles. It suggests using $t$ 6, or $t$ 7 to compensate for multiple obstacles, and precomputing predictor-specific $t$ 8 tables offline indexed by $t$ 9 and $\mathcal{N}_t \subseteq \mathcal{N}$ 0 (Yang et al., 16 Jul 2025). This makes uncertainty handling structurally coupled to predictor selection and active-constraint count.

4. MPC formulation, safety tightening, and theoretical properties

The MPC problem solved by HyPRAP is

$\mathcal{N}_t \subseteq \mathcal{N}$ 1

subject to

$\mathcal{N}_t \subseteq \mathcal{N}$ 2

$\mathcal{N}_t \subseteq \mathcal{N}$ 3

and collision-avoidance constraints of the form

$\mathcal{N}_t \subseteq \mathcal{N}$ 4

with $\mathcal{N}_t \subseteq \mathcal{N}$ 5 and naturally $\mathcal{N}_t \subseteq \mathcal{N}$ 6 (Yang et al., 16 Jul 2025). The constraint function $\mathcal{N}_t \subseteq \mathcal{N}$ 7 is convex and $\mathcal{N}_t \subseteq \mathcal{N}$ 8 is the Lipschitz constant used to inflate the constraint by the conformal radius. Thus prediction uncertainty enters planning as deterministic tightening.

The theoretical analysis studies a simplified two-predictor case with $\mathcal{N}_t \subseteq \mathcal{N}$ 9 more accurate and slower, and $N_t = |\mathcal{N}_t|$ 0 less accurate and faster, satisfying

$N_t = |\mathcal{N}_t|$ 1

If HyPRAP routes $N_t = |\mathcal{N}_t|$ 2 obstacles to $N_t = |\mathcal{N}_t|$ 3 and $N_t = |\mathcal{N}_t|$ 4 obstacles to $N_t = |\mathcal{N}_t|$ 5, the total prediction time is

$N_t = |\mathcal{N}_t|$ 6

hence

$N_t = |\mathcal{N}_t|$ 7

HyPRAP therefore lies between the all-fast and all-accurate single-predictor extremes in prediction-time efficiency (Yang et al., 16 Jul 2025).

For simultaneous safety bounds, the paper defines stacked prediction errors and conformal radii:

$N_t = |\mathcal{N}_t|$ 8

Using Bonferroni’s inequality, it derives

$N_t = |\mathcal{N}_t|$ 9

Under non-interaction, this strengthens to

$P^1=\text{LSTM}$ 0

It also gives a partial-coverage bound requiring all $P^1=\text{LSTM}$ 1 obstacles to satisfy their bounds while only a fraction $P^1=\text{LSTM}$ 2 of $P^1=\text{LSTM}$ 3 obstacles must do so (Yang et al., 16 Jul 2025). These results formalize the intended tradeoff: computation is reduced by routing some obstacles to the cheaper predictor, while strict collective coverage degrades in a controlled way.

The paper’s safety interpretation is conservative MPC rather than direct online optimization of a coherent risk functional. This differs from "Distribution-Free Risk-Aware Planning and Control Under Uncertainty Using Conformal Spectral Risk Control" (Eom et al., 2 Jun 2026), which derives deterministic constraint tightening from conformal calibration of spectral risk measures. This suggests two distinct but compatible directions inside HyPRAP-like systems: predictor-indexed conformal obstacle tubes (Yang et al., 16 Jul 2025) and distribution-free spectral-risk calibration of prediction errors (Eom et al., 2 Jun 2026).

5. Empirical evaluation and quantitative behavior

HyPRAP is evaluated in 1,000 Monte Carlo simulations with 20 to 50 obstacles, using a kinematic differential drive mobile robot, CasADi + IPOPT, and planning/prediction horizons $P^1=\text{LSTM}$ 4 at confidence level $P^1=\text{LSTM}$ 5 (Yang et al., 16 Jul 2025). In an illustrated dense scene with $P^1=\text{LSTM}$ 6 total obstacles, $P^1=\text{LSTM}$ 7 are within sensing range and $P^1=\text{LSTM}$ 8 are routed to active prediction, with high-risk obstacles assigned to LSTM, lower-risk obstacles to GPR, and ignored obstacles rendered gray.

The paper compares three planner architectures: $P^1=\text{LSTM}$ 9 using only LSTM, HyPRAP using hybrid routing, and $P^2=\text{GPR}$ 0 using only GPR. It reports the following navigation results:

Architecture	Success Rate (%)	Avg. Travel Time Step
$P^2=\text{GPR}$ 1	94.0	$P^2=\text{GPR}$ 2
HyPRAP	93.1	$P^2=\text{GPR}$ 3
$P^2=\text{GPR}$ 4	88.5	$P^2=\text{GPR}$ 5

These results place HyPRAP close to the all-LSTM baseline in success rate and travel time, while outperforming the all-GPR baseline, whose larger uncertainty sets produce more conservative trajectories and more deadlocks or infeasibility (Yang et al., 16 Jul 2025). The paper also defines an empirical accuracy measure

$P^2=\text{GPR}$ 6

with

$P^2=\text{GPR}$ 7

The reported average conformal radii are $P^2=\text{GPR}$ 8 for LSTM and $P^2=\text{GPR}$ 9 for GPR, quantifying the accuracy–efficiency asymmetry that motivates routing (Yang et al., 16 Jul 2025).

A separate ablation evaluates P-CRI against proximity-only routing:

Architecture	Success Rate (%)	Total Calls
Prox.A	80.3	587
Prox.B	93.6	1590
P-CRI	93.1	592

At matched compute, P-CRI substantially outperforms Prox.A in safety; at matched success, Prox.B requires roughly two to three times as many predictor calls as P-CRI (Yang et al., 16 Jul 2025). The paper uses this to argue that trajectory-aware relevance scoring is materially better than naive distance heuristics.

A notable empirical observation is that total runtime does not follow prediction time alone. Although $\phi(t,k)$ 0 is fastest in raw prediction, its larger uncertainty sets shrink the MPC feasible region, reduce warm-start effectiveness, and make the nonlinear program harder to solve. This is one of the paper’s main systems-level insights: a uniformly cheap predictor can increase overall solve time by degrading downstream planning geometry.

The explicit HyPRAP formulation in (Yang et al., 16 Jul 2025) sits within a broader family of hybrid prediction–planning systems. In human–robot collaboration, "Anticipate, Adapt, Act: A Hybrid Framework for Task Planning" (Dash et al., 23 Feb 2026) combines next-task prediction with llama-3.3-70b-versatile, symbolic goal grounding, RDDL relational MDP modeling, PROST planning, and reward shaping for preventive and recovery actions. That work does not use the HyPRAP acronym, but it matches a HyPRAP-style pattern in which prediction alters the planner’s objective and stochastic transition structure. Its “risk-aware” component is reward shaping around human action failures rather than formal risk measures.

In automated driving, "Hybrid deep reinforcement learning and planning for safe and comfortable automated driving" (Gupta et al., 2022) uses M2P3 pedestrian path prediction, risk-aware path planning over modified cost maps, hierarchical rule reasoning, and IS-DESPOT/soft actor-critic integration. "HYPE: HYbrid Planning with Ego proposal-conditioned predictions" (Yu et al., 14 Oct 2025) pushes the idea further by conditioning occupancy prediction on each ego proposal and using proposal-guided MCTS refinement. "RAP: Risk-Aware Prediction for Robust Planning" (Nishimura et al., 2022) addresses a different failure mode: under-sampling of long-tail dangerous futures, and therefore shifts the predictive distribution itself to make downstream risk estimation more sample-efficient. These systems suggest that HyPRAP is best understood as an architectural class in which prediction, uncertainty, and planning are coupled at one of several interfaces: goal shaping, cost-map construction, occupancy forecasting, proposal generation, or sample-biasing.

At the same time, the explicit HyPRAP formulation has clear limits. The paper assumes predictor independence in the hybrid conformal section, does not incorporate ignored low-risk obstacles as MPC constraints, and analyzes a simplified two-predictor case in theory (Yang et al., 16 Jul 2025). It also leaves several extensions open: interaction-aware predictors, multi-agent planning, hardware acceleration, and improved routing guarantees. This suggests that current HyPRAP is strongest as a selective-prediction MPC architecture for dense but nonuniformly relevant obstacle sets, rather than a complete theory of risk-aware planning under all forms of uncertainty.

A plausible synthesis is that HyPRAP names a modular design principle rather than a single algorithmic primitive: predict only what matters, quantify predictor-specific uncertainty, and expose that calibrated uncertainty directly to a structured planner. In that sense, (Yang et al., 16 Jul 2025) supplies the explicit acronym and a concrete MPC instantiation, while adjacent work shows compatible variants based on relational MDPs (Dash et al., 23 Feb 2026), POMDP/RL hybrids (Gupta et al., 2022), proposal-conditioned search (Yu et al., 14 Oct 2025), and risk-biased prediction (Nishimura et al., 2022).