Safe Probabilistic Planning for Human-Robot Interaction using Conformal Risk Control
CRC-CBF provides risk-adaptive probabilistic safety for human-robot navigation, but the supplied paper text reports no numerical outcome metrics.
Key Findings
Methodology
The paper introduces CRC-CBF, combining Control Barrier Functions with nonexchangeable Conformal Risk Control. A stochastic human policy predicts actions; predicted barrier values are compared with ground-truth values offline to calibrate a safety margin λ. Online, an LSTM maps interaction features to λ, while a quadratic-program safety filter minimally alters the nominal robot action subject to a probabilistic barrier constraint.
Key Results
- In human-robot navigation simulations, the authors report that the CRC safety filter substantially reduces collisions and safety violations relative to standard CBF and other baselines, while preserving high goal-reaching success and efficient control. However, the supplied text gives no numerical collision rate, success percentage, or runtime, so the magnitude of improvement cannot be independently quantified.
- The main formal result states that if nonexchangeable CRC achieves E[L(λ)]≤α+β, with ε=(α+β)/γ, then enforcing B̂−(λ+ε)≥0 yields Pr(h(xk+1)≥0)≥1−γ. Here β captures total-variation error caused by nonexchangeability.
- Geometric weights wi=ρ^(nk+1−i) emphasize recent interaction history. Higher human-prediction uncertainty increases λ and shrinks the admissible control set; lower uncertainty reduces conservatism while retaining the stated risk guarantee.
Significance
The work addresses a persistent robotics problem: human behavior is multimodal, history-dependent, and difficult to describe with a reliable distribution or worst-case bound. By embedding statistical risk control directly into a CBF constraint, the method links prediction uncertainty to an actionable control buffer and targets per-timestep probabilistic safety. Academically, it connects conformal verification, learned behavior prediction, and invariance-based control. Industrially, it can be attached to existing nominal planners and CBF-QP filters without requiring Gaussian uncertainty models.
Technical Contribution
The paper defines a barrier-error loss L(λ)=max{0,|B−B̂|−λ}, estimates the smallest acceptable margin using geometrically weighted nonexchangeable CRC, and converts an expected-loss guarantee into a high-probability barrier-error bound through Lemma 3. Theorem 1 then establishes Pr(h(xk+1)≥0)≥1−γ. Algorithm 1 performs offline calibration; Algorithm 2 uses LSTM Φ(ϕk) for online margin prediction and solves the CRC safety filter.
Novelty
The authors describe this as the first application of conformal risk control to control-theoretic safety constraints in joint human-robot systems. Unlike conformal trajectory prediction, which primarily calibrates coverage, CRC directly controls a user-defined barrier-error risk. Unlike related conformal decision methods that guarantee only average violation over time, this work derives a single-timestep probabilistic safety statement, while explicitly addressing temporal dependence.
Limitations
- The theory assumes a finite barrier-prediction error bound and depends on the nonexchangeability term β, which is difficult to compute. Rapid distribution shifts, abrupt human maneuvers, or adversarial interactions may therefore weaken practical coverage.
- Experiments are simulation-based human-robot navigation studies. Although the behavior model is trained from real pedestrian data, the current training data do not contain robot actions, leaving real hardware, sensing noise, closed-loop human reactions, and model mismatch insufficiently validated.
Future Work
Important next steps are real-robot and real-pedestrian trials, robot-aware human behavior models, and transparent reporting of collision rates, success rates, latency, calibration curves, and compute cost. Further research could adapt ρ and β online, extend CRC-CBF to multiple robots, and design recovery policies for abrupt distribution shifts or infeasible safety sets.
AI Executive Summary
Autonomous robots operating in hospitals, malls, roads, and warehouses must share space with people whose movements are multimodal and history-dependent. A pedestrian may pass on either side, react to the robot’s previous motion, or change direction abruptly. Gaussian chance constraints simplify this uncertainty, while sampling-based planners can be expensive and generally lack formal guarantees. The central question is whether a robot can remain safe with a stated probability when the human’s next action is unknown.
Gonzales and colleagues propose CRC-CBF, a framework that combines Control Barrier Functions with nonexchangeable Conformal Risk Control. CBFs describe a safe set h(x)≥0 and filter a nominal command through a quadratic program. CRC calibrates the discrepancy between a predicted barrier value B̂ and the ground-truth value B, producing a safety margin λ. Geometrically decayed weights address temporal dependence, and an LSTM predicts λ from current interaction features. Theorem 1 states that, when the CRC risk condition holds and ε=(α+β)/γ, enforcing B̂−(λ+ε)≥0 implies Pr(h(xk+1)≥0)≥1−γ.
In simulated human-robot navigation, the paper reports fewer collisions and safety violations than standard CBF baselines while maintaining high goal-reaching success and efficient control. The supplied text does not provide numerical percentages, dataset names, or runtime measurements, so exact gains cannot be claimed. The conceptual advance is nevertheless clear: uncertainty becomes a calibrated control buffer rather than a fixed worst-case margin. High-risk interactions induce conservative behavior; predictable scenes permit more direct motion. Real-world deployment still requires hardware experiments, complete numerical benchmarks, and robust treatment of rapid distribution shift.
Deep Analysis
Background
CBFs provide forward-invariance guarantees by defining a safe set S={x:h(x)≥0} and constraining the derivative of h. Robust CBFs handle known bounded disturbances, but human actions are often multimodal, history-dependent, and effectively unbounded. Conformal prediction offers distribution-free finite-sample coverage, while Adaptive Conformal Prediction addresses shifts. However, fixed coverage does not directly control the risk of violating a safety constraint, motivating CRC.
Core Problem
Sampling and zero-order-hold control create a gap between continuous-time safety and discrete decisions. Unknown human action uH also prevents the robot from directly solving the deterministic robust CBF-QP. The objective is to choose uR conditioned on history x0:k such that Pr(uR∈Ĉ)≥1−γ, while avoiding excessive conservatism and retaining goal-reaching performance under control and input constraints.
Innovation
- ��First use of nonexchangeable CRC for CBF safety values in joint human-robot systems.
- ��A loss L(λ)=max{0,|B−B̂|−λ} directly measures barrier-prediction error rather than trajectory coverage alone.
- ��Geometric reweighting ρ^(nk+1−i) addresses temporal dependence.
- ��An LSTM maps contextual features to λ, tightening the filter in uncertain interactions and relaxing it in predictable ones.
Methodology
- ��Dynamics: discretize ẋ=f(x)+Bu into xk+1=fd(xk)+BR,duR,k+BH,duH,k.
- ��Robustification: use η=ΔT·Lx(LhLgRu+LhLf+LK∘h) to compensate for sampling error.
- ��Prediction: sample human actions from P(uH|x0:k), integrate predicted states, and compute B̂; use B for offline calibration.
- ��CRC: choose the smallest λ satisfying (ΣwiLi(λ)+B)/(nw+1)≤α.
- ��Filtering: minimize ||uR−unom||² subject to B̂−(λ+ε)≥0.
- ��Online execution: Algorithm 2 computes features ϕk, updates λ=Φ(ϕk), solves the QP, and applies uR,k.
Experiments
The evaluation uses simulated human-robot navigation and crowd scenarios. The human policy is trained on real-world pedestrian data; inputs include agent positions, velocities, and interaction history, and outputs predicted human controls. A nominal robot policy generates offline trajectories. Algorithm 1 uses M trajectories, batches of K, horizon N, and geometric factor ρ to calibrate margins; Algorithm 2 deploys an LSTM online. The paper mentions standard CBF and related baselines, with collision, safety-violation, goal-success, and control-efficiency metrics, but the supplied text gives no dataset names or numerical tables.
Results
The formal result is per-step probabilistic safety: under the CRC risk condition and ε=(α+β)/γ, Pr(h(xk+1)≥0)≥1−γ. Figures qualitatively show larger λ and greater separation in high-uncertainty crowds, and smaller λ in low-risk scenes. Empirically, the authors report fewer collisions and violations while preserving high success and efficient control. Because no numerical values are included in the supplied text, exact baseline deltas and cross-scenario rankings cannot be verified.
Applications
Potential users include service robots, hospital delivery platforms, warehouse vehicles, and autonomous systems navigating crowded public spaces. Deployment requires a human-motion predictor, a computable CBF, a nominal planner, calibration trajectories, and a real-time QP solver. The method is especially relevant when Gaussian assumptions or credible worst-case disturbance bounds are unavailable.
Limitations & Outlook
The guarantees rely on finite barrier error, locally Lipschitz dynamics, feasible controls, and sufficiently gradual distribution shift. β is generally unavailable in closed form, and abrupt behavior may invalidate historical calibration. The current model is trained on pedestrian data without explicit robot actions, potentially missing reciprocal reactions. The supplied paper text also lacks complete numerical benchmarks, hardware validation, and runtime analysis. Future work should address these gaps and test infeasibility recovery.
Plain Language Accessible to non-experts
Imagine a delivery cart moving through a busy shopping mall. Ordinary navigation might look at where each shopper is now and assume they will continue forward. But a shopper can turn left, turn right, stop, or react to the cart, so a single prediction can be wrong at exactly the dangerous moment.
This paper gives the cart a learning safety assistant. The assistant compares past forecasts with what shoppers actually did and measures how wrong its forecasts tend to be. That error becomes an extra buffer around people. When the corridor is crowded and movements are confusing, the assistant asks the cart to keep farther away. When the corridor is open and behavior is predictable, it allows a shorter, more efficient route.
The important idea is not simply “always be extremely cautious.” The assistant changes its caution according to evidence and context. At every step, it chooses the command closest to the cart’s original plan, but blocks commands that are too risky. The paper proves that, under its calibration conditions, the next state remains safe with at least a chosen probability. The supplied text, however, does not report exact percentages, so it supports the method’s direction but not a precise numerical performance claim.
ELI14 Explained like you're 14
Picture a robot in a video game trying to cross a crowded hallway. The other characters are not predictable: one may dodge left, another right, and someone else may react to the robot’s last move. If the robot trusts only one guessed path, crash—game over!
This paper builds a smart safety shield. First, it draws a boundary that the robot should not cross. Then it lets the robot follow its normal plan unless that plan gets too close to danger. The shield also remembers how wrong earlier predictions were. If the crowd looks chaotic, it creates a bigger buffer. If everyone moves predictably, it shrinks the buffer so the robot does not take a ridiculous detour.
The system is called CRC-CBF. CRC acts like a scorekeeper checking prediction mistakes, while CBF acts like an invisible wall. The paper says that if the error-checking step meets its target, the chance that the robot stays safe at the next step is at least 1−γ. Smaller γ means a stricter safety target.
The experiments use simulated human-robot navigation. The authors report fewer crashes and safety violations than ordinary CBF, while the robot still reaches goals efficiently. But the provided paper text does not show exact percentages or real-robot tests. So this is a promising safety shield—not a finished product for every hallway, airport, or hospital yet!
Glossary
Control Barrier Function (CBF)
A mathematical function that represents a safe region and restricts actions so the system does not leave it. Technically, it converts forward-invariance requirements into inequalities that can be enforced by a controller.
The paper constructs barrier certificates and embeds CBF constraints in a quadratic-program safety filter.
Conformal Risk Control (CRC)
A finite-sample statistical method for controlling expected prediction loss with relatively weak distributional assumptions. Unlike ordinary coverage calibration, it can target a user-defined risk function.
CRC calibrates the margin needed to cover errors in predicted safety-barrier values.
Nonexchangeable CRC
A CRC variant for dependent or shifting sequential data. It reweights historical observations because time-series samples are not independent and identically distributed.
The method uses weights wi=ρ^(nk+1−i) for human-robot interaction histories.
Safety margin λ
A buffer subtracted from a predicted safety value to account for prediction error. Larger λ produces a smaller and more conservative admissible control set.
An LSTM predicts λ online from interaction context.
Barrier certificate
A scalar safety value built from system dynamics, the CBF, and a discretization margin. A nonnegative value indicates that a candidate action satisfies the corresponding barrier condition.
The paper distinguishes true B(xk,uk) from computable predicted B̂.
Chance constraint
A constraint requiring a safety event to hold with at least a specified probability rather than for every possible disturbance. It explicitly trades risk against performance.
SRCBF-QP requires Pr(uR∈Ĉ)≥1−γ.
Open Questions Unanswered questions from this research
- 1 The supplied text omits numerical collision rates, success rates, latency, and dataset names; therefore the practical improvement over baselines cannot be quantified.
- 2 The nonexchangeability term β is difficult to compute. Coverage under rapid distribution shift, abrupt human maneuvers, and robot-induced behavioral changes remains unresolved.
- 3 Real hardware, sensor noise, model mismatch, dense crowds, and large-scale multi-agent real-time performance require systematic evaluation.
Applications
Immediate Applications
Hospital and warehouse delivery
A mobile robot can retain an existing goal-reaching planner while CRC-CBF filters commands near people. Required components are a human-motion predictor, geometric CBF, calibration trajectories, and a real-time QP solver. The expected outcome is fewer safety violations without imposing a permanently worst-case buffer.
Public-space service robots
Robots in airports, malls, and campuses can enlarge or shrink separation margins according to crowd density and interaction history. This is useful when pedestrian uncertainty is not credibly Gaussian, but deployment must monitor distribution shift and predictor failures.
Long-term Vision
Verifiable socially aware autonomy
Future systems could combine robot-aware human models, online CRC, and multi-robot CBFs to recalibrate risk throughout long interactions. With standardized real-world benchmarks, the framework could become a safety layer for autonomous machines operating in shared public environments.
Abstract
In this paper, we present a novel probabilistic safe control framework for human-robot interaction that combines control barrier functions (CBFs) with conformal risk control to provide formal safety guarantees while considering complex human behavior. The approach uses conformal risk control to quantify and control the prediction errors in CBF safety values and establishes formal guarantees on the probability of constraint satisfaction during interaction. We introduce an algorithm that dynamically adjusts the safety margins produced by conformal risk control based on the current interaction context. Through experiments on human-robot navigation scenarios, we demonstrate that our approach significantly reduces collision rates and safety violations as compared to baseline methods while maintaining high success rates in goal-reaching tasks and efficient control. The code, simulations, and other supplementary material can be found on the project website: https://jakeagonzales.github.io/crc-cbf-website/.