Biased-MPPI: Robust Trajectory Optimization

• Biased-MPPI is a generalization of MPPI that biases sampling distributions using ancillary controllers, learned policies, and structured priors to enhance trajectory optimization.
• It integrates arbitrary proposal distributions and controller fusion to overcome local minima in nonconvex, safety-critical, and dynamic environments.
• Empirical studies demonstrate that Biased-MPPI improves performance, reduces control effort, and enhances safety through innovative sampling strategies in various robotic applications.

Biased-MPPI refers to a generalization of Model Predictive Path Integral (MPPI) control that informs and shapes the sampling distribution for trajectory optimization. By incorporating information from ancillary controllers, structure priors (such as splines for lane following), learning-based policies, or Control Barrier Functions (CBFs), Biased-MPPI enables efficient, robust receding-horizon control in environments where vanilla Monte Carlo sampling around the previous trajectory fails to provide meaningful exploration. The methodology has been developed and extended in multiple recent works, and variants include controller fusion for arbitrary proposals, neural-network-informed sampling, and CBF-constrained exploration for safety-critical robotics.

1. Foundations: Standard MPPI and Its Limitations

Standard MPPI is an importance-sampling-based receding-horizon stochastic optimal control method. At each control cycle, a set of trajectories is sampled around a nominal control sequence $U$, typically according to a zero-mean Gaussian perturbation $v_t \sim \mathcal{N}(u_t, \Sigma)$. Each sampled control sequence $V$ is rolled out on the true or simulated dynamics, evaluated under a cumulative cost $S(V)$, and assigned an importance weight $\omega(V) = \exp(-S(V)/\lambda)$, where $\lambda$ is a temperature parameter. The nominal control is updated via a weighted average over all samples.

Despite theoretical convergence as the sample count $K$ increases, standard MPPI is highly sensitive to the selection of the sampling distribution. When all samples are drawn around the previous mean—especially in rapidly changing, nonconvex, or cluttered environments—samples may be trapped in high-cost regions, leading to local minima and poor adaptation to new conditions. This vulnerability motivates the deliberate biasing of the sampling distribution to incorporate auxiliary information, escape suboptimal modes, and improve sample-efficiency (Trevisan et al., 2024).

2. Arbitrary Proposal MPPI: Theoretical Generalization

Biased-MPPI extends the path integral framework to permit arbitrary, possibly non-Gaussian, sampling distributions $\mathbb{Q}_s(V)$. The cost function is modified to include a log-ratio correction term,

$$\tilde{S}(V) = S(V) + \lambda \ln \frac{p(V)}{q_s(V)},$$

where $p(V)$ is the original reference distribution, and $v_t \sim \mathcal{N}(u_t, \Sigma)$0 the (possibly multi-modal) biased proposal. The resulting control objective becomes

$v_t \sim \mathcal{N}(u_t, \Sigma)$1

where $v_t \sim \mathcal{N}(u_t, \Sigma)$2 is the controlled distribution and the KL-term quantifies the trade-off between cost minimization and sampler fidelity.

Under this arbitrary-proposal scheme, the optimal importance-sampling weight reduces to

$v_t \sim \mathcal{N}(u_t, \Sigma)$3

i.e., no explicit log-ratio correction appears in the final estimator as long as samples are weighted only by cost (Trevisan et al., 2024).

3. Controller Fusion and Mixture Sampling Strategies

A defining feature of Biased-MPPI is the fusion of multiple ancillary controllers—each likely to generate candidate trajectories corresponding to distinct modes in trajectory space—into the sampler. For example, controller banks may include classical feedback policies (e.g., LQR, LQI), energy-based swing-up for underactuated systems, potential field navigation, learned policies, or structure-driven planners (such as path-following for autonomous driving).

A typical sampling mixture assigns a dedicated number of samples to each ancillary controller trajectory, with remaining samples drawn from a Gaussian centered at the previous solution: $v_t \sim \mathcal{N}(u_t, \Sigma)$4 All $v_t \sim \mathcal{N}(u_t, \Sigma)$5 candidates are scored and weighted via the standard path integral expression (Trevisan et al., 2024). This multi-modal sampling is especially effective in non-convex settings, as it allows the optimization to escape local minima explored by any single controller.

4. Structured Priors and Learned Policies in Biased Sampling

Biased-MPPI can exploit task-structured priors and learned policies to shape its proposal.

• In autonomous driving, spline-based priors parameterize adjacent lane centerlines as cubic Hermite splines, which are tracked by low-level controllers (e.g., Stanley steering, PID). The proposal distribution becomes a mixture, e.g., $v_t \sim \mathcal{N}(u_t, \Sigma)$6, with samples allocated to both prior-tracking and unconstrained branches (Ryu et al., 16 Jul 2025).
• In multi-robot collision avoidance, policies learned through cooperative reinforcement learning (e.g., PPO/IPPO) are embedded as mixture components in the sampler. Samples generated around both standard MPPI and RL seeds are combined, with explicit importance weights incorporating the log-density ratio between proposal and reference (Dergachev et al., 12 Nov 2025).
• In single-step neural-parametric MPPI (Step-MPPI), the proposal at each step is parameterized directly by a neural network policy $v_t \sim \mathcal{N}(u_t, \Sigma)$7, trained over the long-horizon cost with entropy regularization. At inference, rollouts are sampled via $v_t \sim \mathcal{N}(u_t, \Sigma)$8, with learned mean and covariance (Le et al., 2 Apr 2026).

This architectural flexibility enables adaptation to non-stationary environments, dense agent interactions, and highly structured tasks (e.g., driving on structured roads).

5. Barrier-Rate Guided and Safety-Certificate Biased Sampling

Variants of Biased-MPPI address explicit constraint adherence and safety through incorporation of Control Barrier Functions:

• Barrier-Rate MPPI (BR-MPPI) augments the state with rate parameters $v_t \sim \mathcal{N}(u_t, \Sigma)$9, transforms safety inequalities into equality constraints, and projects each sampled control onto the manifold satisfying $V$0 for all safety sets. This projection is solved efficiently in closed form for linearized, control-affine systems. The result is a sampler focused on the invariant set boundary, greatly increasing the probability of safe rollouts (Parwana et al., 8 Jun 2025).
• GPU-Accelerated BR-MPPI applies the same principles in high-dimensional systems (e.g., tractor-trailer with up to nine obstacles), leveraging JAX for GPU-parallelized rollout and constraint projection. Empirically, this yields deterministic collision-free operation at practical rates ($V$1 Hz), outperforming naïve collision-cost-based MPPI (Majd et al., 7 Aug 2025).

These extensions merge forward-invariance safety certification (Nagumo’s condition) with non-myopic sampling, enabling operation closer to constraint boundaries while minimizing computational burden.

6. Empirical Results, Benchmarks, and Observed Benefits

Biased-MPPI has been shown to significantly improve sample efficiency, robustness, and constraint satisfaction across a variety of domains and platforms:

• In rotary inverted pendulum swing-up, median cost was reduced by approximately 25% and control effort by 15% relative to vanilla MPPI or switching among controllers (Trevisan et al., 2024).
• In vessel navigation, fusion with “brake” and “go-slow” ancillary controllers achieved zero collisions across all tested sample counts, and reduced rule violations by half.
• For lane-changing and merging in dense autonomous driving (IANN-MPPI), inclusion of the spline prior improved lane-change success from 75% to 80% and reduced mean merge time by one-third (31.8 s to 21.4 s); costs were similarly reduced (9.49 to 5.90) (Ryu et al., 16 Jul 2025).
• Barrier-Rate MPPI enabled safe quadrotor navigation through narrow corridors where vanilla MPPI could not find feasible paths, even with an order-of-magnitude higher samples (Parwana et al., 8 Jun 2025).
• In articulated vehicle parking, BR-MPPI was the only tested method to reliably avoid collisions, maintaining clearance $V$2, and met real-time constraints with 5,000 rollouts per$V$310 ms step (Majd et al., 7 Aug 2025).
• Step-MPPI achieved a two- to three-fold reduction in per-step latency and up to $V$4 reduction in sample count, while matching or exceeding control performance in robot locomotion, driving, and traffic network management (Le et al., 2 Apr 2026).

A summary of representative experimental results is given below.

Scenario	Vanilla MPPI	Biased-MPPI Variant	Success/Improvement
Pendulum swing-up (Trevisan et al., 2024)	Median cost	–25%	More robust, less effort
AV merging (Ryu et al., 16 Jul 2025)	0.75 success, 31.8 s merge	0.80 success, 21.4 s merge	Faster, higher success, lower cost
Quadrotor corridor (Parwana et al., 8 Jun 2025)	Fails/collides	BR-MPPI succeeds	Reliable planning near constraints
Tractor-trailer (Majd et al., 7 Aug 2025)	Collides	BR-MPPI no collisions	First real-time CBF-certified planning
Step-MPPI (Le et al., 2 Apr 2026)	16.5 ms/step	7.5 ms/step	2$V$5 faster, fewer samples

7. Hyperparameters, Tuning, and Trade-offs

Tuning Biased-MPPI requires balancing the trade-off between fidelity to the proposal and the control cost:

• The inverse temperature $V$6 controls the concentration of the distribution; small $V$7 leads to overfitting to the proposal, while large $V$8 increases weight variance—adaptive normalization is recommended (Trevisan et al., 2024).
• The ancillary mix ratio $V$9, sample count $S(V)$0, sampling covariance, and horizon $S(V)$1 should cover the modes of all auxiliary controllers plus the exploration requirements of the system.
• Proposals induced by strict priors or learned policies can introduce bias if too sharply concentrated, but this bias is always quantifiable via the KL divergence between the proposal and the uncontrolled measure.
• In BR-MPPI, weights $S(V)$2 in the sample projection, barrier buffer widths, and penalty coefficients govern the safety-comfort trade-off and projection accuracy (Parwana et al., 8 Jun 2025).

Conclusion and Perspectives

Biased-MPPI formalizes and implements the fusion of ancillary control strategies, learned behaviors, structural priors, and safety certificates into the MPPI sampling process, yielding a more informative, multi-modal, and task-adaptive trajectory optimization framework. The theoretical foundation ensures convergence to a well-defined stochastic control problem that trades sample efficiency and robustness against quantifiable bias. Empirical evidence demonstrates that Biased-MPPI dramatically improves performance in safety-critical, structured, and rapidly changing settings compared to standard MPPI approaches (Trevisan et al., 2024, Ryu et al., 16 Jul 2025, Majd et al., 7 Aug 2025, Parwana et al., 8 Jun 2025, Dergachev et al., 12 Nov 2025, Le et al., 2 Apr 2026).