DiffCVaR: Reinforcement Learning for Risk Adaptation via Differentiable CVaR Barrier Functions

TL;DR

Learning to adapt risk online is critical for navigation in uncertain crowds. We present an end-to-end risk-adaptive framework that combines reinforcement learning with a differentiable CVaR barrier-function safety layer, jointly learning nominal control, risk level, and safety margin while enforcing probabilistic safety guarantees. This lets the robot stay cautious in difficult interactions without becoming unnecessarily conservative, and it performs strongly across varying obstacle densities, robot models, and OOD crowd shifts.

How Does Risk Adaptation Work?

Risk adaptation combines reactive obstacle-wise risk allocation with a proactive safety margin.

Reactive risk allocation β_i,k: adjusts how much collision risk the policy is willing to tolerate as the perceived danger from obstacle i changes.
Proactive safety margin ΔR_i,k: expands the physical keep-out distance before an obstacle becomes imminent, giving the robot space and time to respond while avoiding unnecessary conservatism.

u_nom

Policy

Produces the nominal control input u_nom,k for task performance.

β_i,k

Reactive Risk Allocation

Reactively allocates the shared risk budget across active obstacles as perceived danger changes.

ΔR_i,k

Proactive Safety Margin

Proactively adjusts the obstacle-wise keep-out distance to create room for maneuver before danger becomes imminent.

Differentiable CVaR-BF-QP Layer

u_nom,k, {β_i,k}, {ΔR_i,k} → Differentiable CVaR-BF-QP → u_k^*

Episode 1

Episode 2

Episode 3

Episode 4

Simulation Settings

Environment & Uncertainty

Workspace: [-6, 6] × [-6, 6] m
Obstacles: 20 dynamic circles; radius 0.4 m; speed in [0.5, 1.5] m/s
Human motion: Social-force policy with avoid_robot = false
Uncertainty: Three-mode GMM; weights [0.6, 0.2, 0.2], standard deviations [0.1, 0.2, 0.2], lateral ratio 0.3

Robot & Control

Dynamics: Single-integrator and unicycle; robot radius 0.3 m
Single-integrator: ||u|| ≤ 1.5 m/s
Unicycle: v ∈ [-1.5, 1.5] m/s; ω ∈ [-1.57, 1.57] rad/s
Simulation: Δt = 0.1 s; 400 steps per episode; goal tolerance 0.6 m
Observation: 20 m sensing range with three-frame history

QP

Policy output: u_nom,k, {β_i,k}, and {ΔR_i,k}
Risk allocation: Total budget 0.5 with β_min = 0.05
Safety margin: Base margin 0.05 m; maximum adaptive increment 1.125 m
QP parameter: Class-K gain α = 2.0
Solver: Differentiable JAX qpth QP layer with slack weight 10.0

Evaluation Protocol

Final statistics 10 base seeds × 50 episodes

Episode horizon 40 s (400 steps)

OOD conditions 30 obstacles and obstacle radius increased to 0.5 m

Compared Methods

We compare optimization-based controllers, learned navigation policies, and hybrid safety-filtered methods.

Optimization

ORCA, CBF-QP, CVaR-BF-QP, and Adaptive-CVaR-BF-QP

Reinforcement Learning

Vanilla RL and CrowdNav++ with constant-velocity or inferred obstacle motion

RL + Optimization

Safety-filtered RL variants, BarrierNet, and the proposed Diff-CVaR-GNN

Statistical Results

Table I baseline comparison for the single-integrator model in a 20-obstacle environment — Table I(a) — Single-integrator

Table I baseline comparison for the unicycle model in a 20-obstacle environment — Table I(b) — Unicycle

Success rate versus obstacle count for the unicycle model — Unicycle

Q1: How do different method types compare in dynamic environments?

RL + Optimization: best safety & efficiency.
Optimization only: lower success rate due to infeasibility.
RL only: efficient but no formal safety guarantee.

Table I

Q2: How do different ways of integrating safety into RL affect performance?

End-to-end training: outperforms decoupled post-hoc safety filtering.
Joint learning eliminates the optimality gap.
Outperforms BarrierNet by modeling stochastic uncertainty.

Table I

Q3: How robust are all methods across obstacle densities?

Proposed method: most robust among others, with graceful degradation.
Obstacle Density: degrade gracefully.

Figure I

Table II out-of-distribution generalization for the unicycle model. Case I uses 30 obstacles; Case II increases the obstacle radius to 0.5 meters — Table II — Unicycle
**Case I:** high obstacle density (30 obstacles). **Case II:** increased obstacle radius (0.5 m).

Q4: How robust are different type of methods under OOD environment changes?

Proposed method: adaptation gives best OOD success rate.
RL only: degrades most under unseen policy shifts.
RL + Safety Filter: regains robustness but still lags behind.

Table II

GMM Setting	Weights [w₁, w₂, w₃]	Stds [σ₁, σ₂, σ₃]	Lat.
Matched / oracle GMM	[0.6, 0.2, 0.2]	[0.1, 0.2, 0.2]	0.3
A.1 Policy overestimates uncertainty	[0.8, 0.1, 0.1]	[0.05, 0.1, 0.1]	0.2
A.2 Policy underestimates uncertainty	[0.4, 0.3, 0.3]	[0.2, 0.4, 0.4]	0.4
B Imbalanced mode weights	[0.9, 0.05, 0.05]	[0.1, 0.2, 0.2]	0.3

GMM parameter settings. The policy keeps the matched/oracle belief fixed while the environment parameters change. Lat. denotes the lateral-mode displacement ratio.

GMM trajectory examples

All animations use the same reset seed and a fixed policy; only the environment GMM changes. The matched/oracle animation is the common baseline for both groups.

Group A — uncertainty mismatch

Six-second environment trajectory animation for the matched oracle GMM — **A.0** Matched / oracle GMM

Six-second environment trajectory animation when the policy overestimates GMM uncertainty — **A.1** Policy overestimates uncertainty

Six-second environment trajectory animation when the policy underestimates GMM uncertainty — **A.2** Policy underestimates uncertainty

Group B — mode-weight mismatch

Six-second environment trajectory animation for the matched oracle GMM baseline in Group B — **B.0** Matched / oracle GMM

Six-second environment trajectory animation for imbalanced GMM mode weights — **B.1** Imbalanced mode weights

GMM Setting	Success Rate (%)	Collision Rate (%)
Matched / oracle GMM	97.0	2.6
A.1 Policy overestimates uncertainty	97.4	2.4
A.2 Policy underestimates uncertainty	92.4	6.2
B Imbalanced mode weights	96.8	2.6

Table III — Unicycle

Q5: How robust is the method to GMM mismatch?

The method shows empirical robustness to moderate GMM mismatch, with graceful degradation when uncertainty is underestimated.

Table III

Ablation study of adaptive risk allocation, adaptive safety margin, and the safety layer for the unicycle model — Table IV — Unicycle
Policy 1: full; Policy 2: fixed risk; Policy 3: fixed margin; Policy 4: both fixed; Policy 5: no safety layer.

Q6: How does each adaptive component affect performance?

Each adaptive component contributes: the full method achieves 97.0% SR and 2.6% CR, while removing the safety layer causes the largest degradation.

Table IV

Citation

@article{wang2026diffcvar,
  author  = {Xinyi Wang and Taekyung Kim and Bardh Hoxha and Georgios Fainekos and Dimitra Panagou},
  title   = {{DiffCVaR}: Reinforcement Learning for Risk Adaptation via Differentiable {CVaR} Barrier Functions},
  journal = {IEEE Robotics and Automation Letters},
  pages   = {1--8},
  year    = {2026},
  doi     = {10.1109/LRA.2026.3742268}
}