Robots that navigate with potential fields (a virtual pull toward the goal plus virtual pushes away from obstacles) get stuck wherever the pull and the pushes cancel. The usual fixes all go around the obstacle. One fix follows the wall. Another rotates the push so the robot slides along the boundary. A third samples several candidate paths so the planner can choose a different gap. These fixes share one hidden assumption: every boundary is rigid and must never be touched. Real clutter is often not like that. Curtains, hanging cables, plastic strip doors, foliage, light boxes and swinging flaps can all be nudged aside with a few newtons of force. For such obstacles, a gentle push is usually shorter and safer than a long detour.
We propose a planner that treats "push through gently" as one more candidate route, alongside "go around," inside a single probabilistic planner. Pushing is allowed only when the robot has evidence that the obstacle yields. The force allowed shrinks automatically as the robot gets closer to things it cannot identify. The robot gathers that evidence with a technique borrowed from wireless engineering. Reflecting surfaces in wireless networks cannot sense the signals they reflect, yet they can be tuned by trying random configurations and averaging a single scalar measurement. A robot pressed against an unknown obstacle is in a similar position: it sees only a scalar force residual. We let it probe in random directions and keep running averages of how the obstacle responds, rather than modeling the contact physics.
Core Claim
Our claim is that potential-field deadlocks in cluttered spaces with mixed rigid and compliant obstacles are better resolved by inference over two things jointly: which way to go around, and whether to push through. A detour-only method cannot do this, however clever it is. This requires three ingredients that no paper in our file combines.
The first ingredient is the planner. Control as inference recasts planning as computing a probability distribution over action sequences that are likely to be "optimal," with optimality likelihood $\exp(-\text{cost})$ (Levine et al.). Stein variational model predictive control (SV-MPC) approximates this distribution with a small set of trajectory "particles." A kernel term pushes the particles apart so that they keep distinct routes, such as left versus right of an obstacle (Stein Variational Model Predictive Control). We reuse this machinery unchanged.
The second ingredient is the contact term. We add a likelihood term built on dynamic contact thresholding, the idea from Contact Anticipation for pHRI that the force level counted as acceptable contact should fall as proximity to unknown surfaces grows. In that paper, the falling threshold sits inside a single-step reactive quadratic program. In our planner, it becomes a cost that can make contact-making particles likely:
$$ p(O_t = 1 \mid \mathbf{x}_t, \mathbf{u}_t, \theta) = \exp\!\Big(-c_{\text{task}}(\mathbf{x}_t,\mathbf{u}_t) - \lambda\,\big[\hat f_c(\mathbf{x}_t,\mathbf{u}_t;\theta) - \tau(d_t, \sigma_\theta)\big]_+\Big) $$
The symbols are as follows. $O_t$ is the binary optimality variable at step $t$, $\mathbf{x}_t$ and $\mathbf{u}_t$ are state and action, and $c_{\text{task}}$ is the usual goal and effort cost. $\hat f_c$ is the predicted contact force under the current estimate $\theta$ of local obstacle compliance. $\tau(d_t,\sigma_\theta)$ is the allowed force: it falls with proximity $d_t$ to unclassified surfaces and with the uncertainty $\sigma_\theta$ of the compliance estimate. $[\cdot]_+$ keeps only the positive part, and $\lambda$ is a penalty weight. A rigid obstacle (high stiffness) or an uncertain one (large $\sigma_\theta$) drives $\tau$ toward zero, which recovers ordinary avoidance. Known-soft obstacles open a corridor through the barrier.
The third ingredient is how $\theta$ is estimated. We model this on blind passive beamforming. That work's Conditional Sample Mean method applies random configurations to a sensing-blind reflecting surface, records received power, and averages conditionally per element and state to find a near-optimal configuration with no channel model (Blind Passive Beamforming for MIMO System). Our robot plays the same game. When the planner detects a stall, the robot applies short, low-force probe pushes in randomly drawn directions within its tangent cone. It records only the scalar force residual and displacement, then forms per-direction conditional means of yield-per-newton. These means feed $\theta$ and $\sigma_\theta$.
The analogy is not decorative. The joint torque residual is a product of unknowns (contact location, contact normal, stiffness). This is structurally like the cascaded RIS channel (Channel Estimation for RIS-Empowered Multi-User MISO), which is the end-to-end composite of several unknown links. In both cases, scalar-feedback averaging sidesteps the factorization.
Research Program
We build three components.
The first is a contact-aware SV-MPC. It keeps the structured Markovian kernels of SV-MPC with 16–32 particles. It adds a lightweight contact predictor $\hat f_c$, a spring along the contact normal with stiffness drawn from the current posterior over $\theta$, and the threshold function $\tau$. The repulsive kernel matters especially here. Without it, the "push" particles and "go around" particles would average into a trajectory that half-pushes and half-detours. This is exactly the failure that Lambert et al. document for MPPI, where averaging competing routes drives the robot into the central barrier.
The second is a stall-triggered probing module. It runs only when the planner's expected progress drops below a threshold, so free-space behavior is unchanged. It uses a probe budget of $T$ pushes, swept from 4 to 32, and returns $\theta$ and $\sigma_\theta$. The force cap during probing is fixed at the lowest value of $\tau$. This guards against the energy-injection risk that rapidly varying thresholds pose for the passivity results of Huber et al.. We also adopt their state-dependent damping during any contact phase.
The third is an analysis of when pushing should win. A detour costs roughly the extra path length around the obstacle's perimeter. A push costs probing time plus a contact penalty. Comparing the two predicts a crossover: pushing wins when barriers are long relative to their thickness and are truly compliant. We test this prediction directly rather than assume it.
The scope is narrow on purpose. Humans are always classified as unknown, so $\tau$ goes to its floor near them. The work concerns objects, not physical human-robot interaction.
Baselines
The naive reference is a standard artificial potential field with a fixed conservative contact threshold and emergency stop. On symmetric and U-shaped traps, such methods (along with ORCA, μNAV and RPF) succeed below 15% of the time according to Escaping Local Minima (APF-WF). We expect the same on our rigid variants and run it to anchor the scale.
The strongest detour baseline is APF with wall-following, in both its rule-based and learned (Vision Transformer) switch variants, from the same paper. It reports up to 100% success and at least 94.2% arrival on its four environments. Its weakness, admitted by its authors, is inflated makespan. The file gives no makespan figure, so we report "no reported result found on Lacuna" for that metric and measure it ourselves. We use the authors' 1,800-trajectory demonstration set for the learned switch. To keep the comparison fair, we give it the same compliant-obstacle environments. We also give it the advantage of an oracle that lets it ignore obstacles labeled soft, which isolates whether inferred compliance is needed.
The strongest inference baseline is SV-MPC itself, with obstacles as soft costs. It reports 96% success versus about 64% for MPPI and CEM in planar navigation, and an average cost of 411.26 versus 813.13 for MPPI on the Franka narrow-slot task (Stein Variational Model Predictive Control). We reproduce both numbers before adding obstacles of mixed compliance. All MPC variants share the particle count, horizon, iteration count per control step, and a hyperparameter search of 50 trials each.
Two ablations isolate the key ingredients. The first is contact-aware MPPI, our likelihood with a unimodal Gaussian planner, which tests whether the multi-route posterior is needed. The second is our method with a fixed $\tau$ and no probing, which tests whether the threshold must depend on proximity and evidence.
Beating the field means two things. On rigid-only environments we must match APF-WF and SV-MPC success within confidence intervals. On mixed environments we must cut makespan by a clear margin, which we set in advance at 25%. We must achieve this without raising peak contact force on rigid obstacles above the fixed-threshold baseline, and with a false-push rate on rigid obstacles under 2%.
Why not just…?
Why not just run SV-MPC with a soft obstacle cost, since soft costs already permit penetration? This is the most serious threat. A soft cost lets particles cut through any obstacle in proportion to the cost weight, rigid and compliant alike. It therefore must either be tuned conservatively, losing the shortcut, or aggressively, hitting walls. The fixed-$\tau$ ablation and the soft-cost SV-MPC baseline settle this. If either matches our makespan with a comparable rigid false-push rate, the probing ingredient is unnecessary, and we would report that.
Why not just use wall-following, which already reaches 100% success? Success is not the problem; detour length is. The APF-WF paper itself flags makespan inflation. The oracle-compliance variant described above tells us how much a detour planner would gain from perfect knowledge, which bounds what our inference can add.
Why not just classify compliance from vision instead of probing? Visual appearance poorly separates a stiff curtain from a slack one, or a loaded box from an empty one. Probing measures yield directly. Still, this is a fair alternative. We add a variant in which an oracle visual label sets $\theta$ with fixed $\sigma_\theta$, and we report the gap. If the gap is small, probing is only worth it for visually ambiguous objects.
Why not just add the contact cost to MPPI? The contact-aware MPPI ablation answers this. We predict it will blend pushing and detouring particles into collisions, as MPPI does with left/right routes in Lambert et al.
Proposed Experiments
We extend two existing benchmarks. In the four APF-WF environments (three guided robots, three uncontrolled ones), we replace 30–70% of the barrier segments in U-traps and corridors with compliant elements of varying stiffness. Examples include hinged flaps, hanging strips and pushable boxes, simulated in PyBullet. We also extend the Franka narrow-slot task of SV-MPC with a compliant flap across one of two routes.
Metrics are success, makespan, mean timesteps, peak and integrated contact force, rigid false-push rate, and probe time. No reported result was found on Lacuna for any method on compliant-obstacle variants, so all numbers on those variants are new. A hardware check on the ToF-instrumented Franka testbed of Contact Anticipation for pHRI tests whether probing keeps impact forces at or below the roughly twofold reduction that paper reports versus non-anticipatory control.
Three outcomes decide the proposal. First, as the fraction of compliant segments rises, the makespan advantage over APF-WF should grow while rigid false-push stays flat; if it does not grow, the idea fails. Second, on rigid-only maps, our method should reduce exactly to SV-MPC performance. Third, error in the compliance estimate should fall with probe budget $T$ at roughly the $1/\sqrt{T}$ rate expected of a sample mean, as in the conditional-mean analysis of blind beamforming. If error plateaus well above zero, the scalar-feedback analogy breaks, and a model-based contact estimator should replace it.
References
- Variational Policy Search via Trajectory Optimization
- Stein Variational Model Predictive Control
- Contact Anticipation for Physical Human-Robot Interaction with Robotic Manipulators using Onboard Proximity Sensors
- Blind Passive Beamforming for MIMO System
- Channel Estimation for RIS-Empowered Multi-User MISO Wireless Communications
- Passive Obstacle-Aware Control to Follow Desired Velocities
- Escaping Local Minima: Hybrid Artificial Potential Field with Wall-Follower for Decentralized Multi-Robot Navigation