acceptodds
Under review as a conference paper at ICLR 2027

Taming the Learning Dynamics of Neural Barriers Using Projected Anchor Sets

Abstract

Learning neural control barrier functions (CBFs) requires shaping safety boundaries of complex high-dimensional geometry while finding safety-preserving controls along them. Direct training of neural network representations couples neural network training, worst-case state search, and control selection in a difficult min–max–min optimization problem that can destabilize learning. We propose new approaches for taming the learning dynamics of neural barriers, grounded in Nagumo’s theorem that characterizes forward invariance through boundary conditions. Our formulation exploits a common structure in safety problems: violations are observable in a low-dimensional space even when the system state is high-dimensional. Our method, projected anchor sets, uses these observations to shape a barrier in the full state space. Multistep recovery searches backtrack along failed rollouts to states where changing the actions can still avert failure. These states anchor the candidate zero level set, while their recovery trajectories supervise both barrier fitting and policy fine-tuning. The resulting loop learns from its own safety feedback, without expert safety demonstrations or a separate task reward. We connect the formulation to high-order CBFs and specify conditions under which the learned boundary and policy ensure forward invariance.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Neural Control Barrier Functions, Safety, Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_1gprOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

NBF-PAS trains neural control barrier functions without external safety oracles. It backtracks along failed rollouts, searching alternative actions for recoverable full-state "anchors" guided by low-dimensional safety geometry. Anchors and recovery trajectories supervise barrier fitting and policy fine-tuning, improving boundaries and collision avoidance on Dubins, Segway, quadrotor, and Hopper systems.

Critical strengths

The paper takes on a real bottleneck in neural CBF learning: existing methods depend on expert demonstrations, safe policies, or hand-designed reward signals. NBF-PAS gets its supervision from the system's own counterfactual recoveries. Backtracking from failures to full states where some alternative action still avoids collision gives anchors that fix both where the zero level set sits and which way its normal points.

R_norm shows baselines with low collision rates still have poorly oriented normals (Dubins 9.9% vs. 3.6%; quadrotor 22.3% vs. 0.0%). it directly supports the claim that anchors fix what sign and feasibility losses leave underdetermined.

Critical weaknesses

There is only one baseline, can compare with safe RL methods? The idea is good but evaluation is too thin: one baseline, no variance reporting, and only single-obstacle setups. Safety gains partly come from conservatism (100% falls on Hopper before retraining).

The paper has a strong assumption, requires resettable counterfactual rollouts plus a hand-specified projection k and margin g, which effectively means knowing the obstacle geometry.

Questions for authors

There is only one baseline, can compare with safe RL methods?

1 comment

Weakness 1-1 (CPO baseline with variance reporting):

We added a new CPO baseline initialized from scratch and from nominal controller weights. We did not include safe RL methods because our proposed NBF-PAS focuses on learning a neural barrier that captures a more correct landscape compared to conventional training, instead of establishing our superiority over all safe RL methods.

While safe RL methods are powerful tools in collision avoidance tasks, if the solution is too narrow and the goal reward is far, RL methods maybe sensitive to seeds if exploration cannot find the narrow solution. If RL must update the policy given an existing nominal one, it may get trapped at negotiating between staying close to the nominal policy and improving long-term discounted reward. Segway collision avoidance is such a challenging task that successful passage requires coordinated tilting and forward motion. An easy local minimum is just to stand still in front of the obstacle. We use this as the example to show why our learned neural barrier can help in Appendix D.4 with four seeds.

MethodsSuccess RateCollision RateFall Rate
NBF-PAS98.55% ± 0.580.0%0.5%±0.61
CPO initialized from scratch initialization25.33% ± 29.7622.25% ± 25.3649.42% ± 57.07
CPO initialized with nominal controller0.00%±0.0097.58% ± 0.170% ± 0

Among four seeds (40-43), only seed 41 and 42 that initialized from scratch initialization are able to improve with 2M training steps. Collision rates are extremely high when CPO is initialized with the nominal controller that is not trained with collision avoidance.

Weakness 1-2 (Multiple-obstacles Setting):

We added a two-obstacle Segway scenario (Appendix D.5) to show that our neural barrier trained with one obstacle environment can be directly used as the auxiliary loss in PPO training for two-obstacle environments. The two obstacles have the same height and radius as the training obstacle but different horizontal positions. The barrier network BθB_{\theta} is kept frozen throughout; only the control policy is adapted. Since Because the Segway dynamics are invariant to horizontal translation, we evaluate the same network in coordinates relative to each obstacle:

Bi(x)=Bθ(Tix),Tix=(p−pi,ϕ,p˙,ϕ˙),Bmulti(x)=min⁡i=1,2Bi(x)B_i(x)=B_\theta(T_i x), T_i x=(p-p_i,\phi,\dot p,\dot\phi),B_{\mathrm{multi}}(x)=\min_{i=1,2}B_i(x)

The policy is updated through

Lchange=Ex∼Dactive[[−ΔBmulti−Δt,γ(Bmulti(x)−ℓ)]+2]\mathcal{L}_{\mathrm{change}}=\mathbb{E}_{x\sim\mathcal{D}_{\mathrm{active}}}\left[\left[-\Delta B_{\mathrm{multi}}-\Delta t\\,\gamma(B_{\mathrm{multi}}(x)-\ell)\right]_{+}^{2}\right]

Warm-started single-obstacle controller, we are able to obtain 98.13%±1.41 success rate with only 1.5%±1.64 collision rate across four seeds.

Weakness 1-3 (Falling):

We would like to clarify that the falling of Hopper does not falsify the learning of neural barrier itself. Before policy adaptation, the barrier filter modifies the nominal action to satisfy the learned obstacle-avoidance constraint while minimizing action deviation. The criterion does not preserve gait stability. Even a small action modification may drive the Hopper into a state where nominal policy has never seen and knows how to react. The pre-retraining result therefore shows that in practice, direct filtering alone is insufficient for successful Hopper control, and this is why in practice most paper also formally proposes to iterate on the policy update like PNCBF. The way we propose to train a CBF can actually replace how CBF is learned while combining the policy update section.

Weakness 2:

We do assume access to geometric safety specification and a resettable simulator that allows restoration of sampled states and transition queries under alternative actions. For neural CBF learning, sign supervision on safe and unsafe states must exist to help place the learned nonnegative set within the physically admissible region as explained in Appendix A. Our projection κ\kappa and margin gg provide inexpensive geometric labels without requiring expert annotations from a given safe controller. This access supports both counterfactual rollouts for boundary supervision and the local action-affine dynamics approximation used in policy refinement. However, we do not require analytical dynamics derivatives.

Question:

We added CPO as an additional baseline in Appendix D.4 to show why our learned neural barrier can help with four seeds.

Reviewer_QPGYOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes NBF-PAS, which learns neural barrier functions from recovery trajectories found by backtracking along failed or near-collision rollouts. These samples guide the location and orientation of the safety boundary. Experiments on four simulated systems show lower collision rates, while additional policy training helps recover task performance on Hopper.

Critical strengths
  1. The paper makes a useful distinction between being collision-free and still being able to avoid a collision. This helps explain why the safety boundary needs to account for velocity, orientation, and control limits.
  2. I like the idea of using recovery searches to generate supervision without expert safe actions. Searching through a low-dimensional projection while learning the barrier from the full state is a sensible choice.
  3. The Segway result is encouraging: the fall rate drops from 100% to 3.1% while collisions remain at zero. The boundary plots and capacity study also give some intuition for what the method learns.
Critical weaknesses
  1. I am not fully convinced by the step from finding a recovery to placing a boundary anchor. Passing the local recovery test does not necessarily mean the trajectory will remain safe. Even if recovery succeeds, why should the selected state lie on the zero level set rather than inside the safe set?
  2. The comparisons are somewhat limited. PNCBF would be a useful baseline, and separate ablations of anchor placement, normal alignment, and iterative search would help show what drives the improvement. The main results would also benefit from reporting variation across seeds.
  3. The Hopper result is harder to interpret. Before policy retraining, the method avoids all collisions but falls in every rollout. Retraining helps considerably, though it leaves open how much the barrier itself contributes. A policy trained directly on the same recovery data would be a useful comparison.
Questions for authors
  1. Which anchor loss was used? Section 4.3 and Appendix F use projected gradients, but Appendix E questions this formulation. Could you clarify the difference?
  2. With negative recovery thresholds, an accepted action can still move toward the obstacle. Do you check whether the following trajectory remains collision-free?
  3. How do you compute policy gradients through the auxiliary loss with a black-box simulator, particularly for Hopper?
4 comments

Weakness 1:

Thank you for the point. We agree that our proposed anchors construction does not imply long-horizon safety invariance. However, these anchors remain useful because they provide soft geometric targets around which the barrier’s invariance conditions are fitted.

Each anchor specifies a preferred boundary location through placement penalty min⁡θBθ(xt+1)2\min_{\theta}B_{\theta}(\mathbf{x}_{t+1})^2, not a state that the learned safe set must include, while normal alignment min⁡θ(n^(xt+1)⊤nθ(xt+1))2\min_{\theta}(\hat{n}(\mathbf{x}_{t+1})^{\top}n_{\theta}(\mathbf{x}_{t+1}))^2 supplies a preferred local orientation. Importantly, both are soft penalties optimized jointly with the full-state CBF conditions, so training seeks a boundary that reconciles the anchor’s geometric guidance with the dynamical conditions for invariance. Recovery anchors thus gives the learner explicit geometric targets beyond the unsafe-state labels supplied by terminal failures.

Weakness 2-1 (Comparison with PNCBF):

Thank you for suggesting PNCBF. We have added this baseline in Appendix D.3 and evaluated both methods as safety filters for unseen MPC on 400 held-out initial state-goal pairs. As reported below (Table 6 in the revised paper), NBF-PAS achieves higher success rate than PN-CBF with zero collision rate.

MethodsSuccess rateCollision rate
MPC81.75%8.75%
NBF-PAS90.50%0%
PNCBF stage 1 + MPC88.75%0%
PNCBF stage 3 + MPC alone82.75%7.75%
PNCBF stage 3 + nested QP87.25%1.50%

The fundamental distinction lies in how the barrier is learned. PNCBF learns a a barrier by estimating which states remain safe under its current controller at a given goal, then repeat this process after improving the controller through QP-based filtering. NBF-PAS directly supervises the barrier’s boundary placement and orientation through recovery anchors grounded in physical obstacle geometry and realizable control responses. This gives barrier learning an explicit geometric reference beyond the behavior of the current controller. Figure 13 in the revised paper shows the resulting difference in the learned geometry. NBF-PBS learns heading-dependent unsafe regions that consistently enclose the physical obstacle, independently of the controller’s behavior. In contrast, PNCBF initially labels collision-irrelevant states far from the obstacle as unsafe (stage 1), as Bellman updates propagate future collision risk backward along failure trajectories. After iterative controller improvement and barrier retraining (stage 3), these spurious unsafe regions disappear, but even the physical obstacle is classified as safe. Thus, PNCBF’s learned geometry reflects the evolving controller’s future risk rather than the underlying physical safety constraints, whereas NBF-PAS preserves obstacle-consistent safety boundaries.

Weakness 2-2 (Isolation on anchor placement and normal alignment):

We added new ablation in Appendix D.2 to examine the contribution of this anchor supervision. Removing both placement and alignment disables the use of anchor states, while retaining the standard neural CBF fitting. The resulting model assigns negative barrier values throughout the 3D geometry, whereas the full objective retains a localized, closed boundary. This illustrates the value of anchor supervision: it helps preserve boundary structure that minimizing the standard empirical losses through training demonstrations alone does not ensure. The table below shows the collision rate, barrier violation rate, and unsafe area across 5 seeds in Quadrotor system. Specifically, barrier violation rate computes the rate of states that fail the CBF condition among visited states, and unsafe area is the fraction of state-space grid area where B(x)<0B(x)<0 for all visited states.

SettingsCollision RateBarrier Violation RateUnsafe Area
NBF-PAS1.9% ± 1.113.2% ± 3.226.6% ± 7.0
No Anchors14.8% ± 3.39.8% ± 3.944.2% ± 18.3
No Normal Alignments6.4% ± 4.411.9% ± 4.835.0% ± 6.2
No Both26.9% ± 1.013.2% ± 2.668.9% ± 3.6
Baseline27.1% ± 8.026.7% ± 4.966.3% ± 17.6

Recall that the baseline minimizes the ∇Bθ⊤x˙+γBθ\nabla B_{\theta}^{\top}\dot{\mathbf{x}}+\gamma B_{\theta} while no both only minimizes ∇Bθ⊤x˙\nabla B_{\theta}^{\top}\dot{\mathbf{x}}. Removing anchor placement increases the collision rate from 1.9% to 14.8% and the fraction of the evaluated region labeled unsafe from 26.6% to 44.2%. Removing normal alignment increases collisions to 6.4% and the unsafe-labeled fraction to 35.0%. Removing both produces 26.9% collisions and labels 68.9% of the evaluated region unsafe—approximately 2.6 times the fraction for NBF-PAS.

Iterative search helps because the training losses are reducing per iteration as shown in Figure 10, and the quality of how neural barrier memorizes the recovery anchors is determined by these empirical losses. Meanwhile, collision rates during evaluation also decreases initially, with mild oscillation when fitting to a smaller loss.

Weakness 3:

Thank you for raising this distinction. RL retraining is intended to highlight that the learned barrier remains useful for adapting policies not used during barrier training, without retraining the barrier itself. In the simpler systems, the original nominal controller can still complete the task under barrier filtering. However, Hopper’s contact dynamics and balance requirements make locomotion sensitive to control changes: filtering can shift states outside the nominal policy’s training distribution and disrupt its learned gait, motivating policy adaptation. To isolate the barrier’s contribution, we include a pure RL baseline trained without barrier guidance under matched training conditions (Table 2). This is a PPO controller with collision penalty in rewards. Barrier-guided adaptation achieves higher goal-reaching success and fewer collisions, demonstrating that the fixed barrier improves safety and task completion beyond barrier retraining. We will clarify this implication in the revision.

Besides, we added another safe RL baseline CPO that initialized from scratch and from nominal controller weights and compared it with our NBF-PAS. Segway collision avoidance is a challenging task that successful passage requires coordinated tilting and forward motion. An easy local minimum is just to stand still in front of the obstacle. We use this as the example to show why our learned neural barrier can help in Appendix D.4 with four seeds.

MethodsSuccess RateCollision RateFall Rate
NBF-PAS98.55% ± 0.580.0%0.5%±0.61
CPO initialized from scratch initialization25.33% ± 29.7622.25% ± 25.3649.42% ± 57.07
CPO initialized with nominal controller0.00%±0.0097.58% ± 0.170% ± 0

We would like to clarify that the falling of Hopper does not falsify the learning of neural barrier itself. Before policy adaptation, the barrier filter modifies the nominal action to satisfy the learned obstacle-avoidance constraint while minimizing action deviation. The criterion does not preserve gait stability. Even a small action modification may drive the Hopper into a state where nominal policy has never seen and knows how to react. The pre-retraining result therefore shows that in practice, direct filtering alone is insufficient for successful Hopper control, and this is why in practice most paper also formally proposes to iterate on the policy update like PNCBF. The way we propose to train a CBF can actually replace how CBF is learned while combining the policy update section.

Meanwhile, our pure-RL comparison already isolates the effect of the having the neural barrier in the training. As specified in Appendix B.4, both methods use the same nominal-policy initialization, environment, reward, interaction budget, and nominal-action imitation loss. NBF-PAS additionally uses the barrier change auxiliary loss, which guides policy updates toward actions that preserve the learned barrier at currently visited safety-critical states. Thus, the comparison assesses the contribution of barrier-based guidance beyond policy retraining alone.

Question 1:

The experiments use the anchor loss given in Section 4.3 and Appendix F, comprising a zero-level placement term and two projected alignment terms.

Lanchor(θ)=E(xt,xt+1)∼Danchor[Bθ(xt+1)2+max⁡(−n^(xt+1)⊤nθ(xt+1),0)2+max⁡(−v(xt,xt+1)⊤nθ(xt+1),0)2]\mathcal{L}_{\mathrm{anchor}}(\theta)=\mathbb{E}_{(\mathbf{x}_{t},\mathbf{x}_{t+1})\sim\mathcal{D}_{\mathrm{anchor}}}\left[B_{\theta}(\mathbf{x}_{t+1})^2+\max\left(-\hat{\mathbf{n}}(\mathbf{x}_{t+1})^{\top}\mathbf{n}_{\theta}(\mathbf{x}_{t+1}),0\right)^2+\max\left(-\mathbf{v}(\mathbf{x}_{t},\mathbf{x}_{t+1})^{\top}\mathbf{n}_{\theta}(\mathbf{x}_{t+1}),0\right)^2\right]

To define this one-way trajectory precisely, we use the Jacobian-based mapping to compute projection (Appendix E). Let Jκ(x)=Dκ(x)J_\kappa(x)=D\kappa(x). We use the workspace alignment vector qλ(x)=(Jκ(x)Jκ(x)⊤+λI)−1Jκ(x)∇xBθ(x)q_\lambda(x) = \bigl(J_\kappa(x)J_\kappa(x)^\top+\lambda I\bigr)^{-1}J_\kappa(x)\nabla_x B_\theta(x) that solves the regularized least-squares problem \min_{q}\\;\left\\|J_{\kappa}(x)^{\top}q-\nabla_{x}B_{\theta}(x)\right\\|_{2}^{2}+\lambda\\|q\\|_{2}^{2}.. That is, Jκ(x)⊤qJ_{\kappa}(x)^{\top}q is the approximation of full-state barrier gradient. We use Jacobian because it captures how multiple configuration variables jointly move the safety-relevant body part.

Question 2:

We allow a negative threshold because even trajectories on a true invariant safety boundary can move toward an obstacle in workspace. For example, in Appendix F, we show a state on the numerical Hamilton-Jacobi Reachability zero contour whose velocity points partly toward the obstacle (n^⊤v^=−0.607\hat n^\top\hat v=-0.607). Negative projected alignment therefore does not imply crossing the safety boundary: invariance depends on the joint evolution of position and velocity. Requiring non-negative alignment would exclude such valid boundary transitions. Allowing ρ<0\rho < 0 keeps these candidates eligible as soft geometric targets, while the full CBF conditions remain unchanged. We do not check if the following trajectory remains safe, because potential future collisions can be captured by iterative training. When neural network learns to fit the barrier by current violation, new collisions will undergo the same backtracking search and provide new anchor data to the training.

Question 3:

We approximate the policy gradient by querying MuJoCo to construct a local action-affine approximation of one-step transition. Specifically,

f^(x)=FΔ(x,0)−xΔ,G^:,j(x)=FΔ(x,ϵej)−FΔ(x,−ϵej)2ϵΔ\hat f(x) = \frac{F_\Delta(x,0)-x}{\Delta}, \qquad \hat G_{:,j}(x) = \frac{F_\Delta(x,\epsilon e_j)-F_\Delta(x,-\epsilon e_j)} {2\epsilon\Delta}

where eje_j is the jjth standard basis vector in the action space. We then evaluate the auxilary barrier change loss using

x^ϕ+=x+Δ(f^(x)+G^(x)μϕ(x))\hat{x}^{+}_{\phi}=x+\Delta\bigl(\hat{f}(x)+\hat{G}(x)\mu_{\phi}(x)\bigr)

During this update, the sampled state and estimated dynamics coefficients are held fixed. Automatic differentiation propagates through the policy, this surrogate transition, and the barrier network with respect to the input as

∇ϕBθ(x^ϕ+)=(Dϕμϕ(x))⊤ΔG^(x)⊤∇xBθ(x^ϕ+)\nabla_{\phi}B_{\theta}(\hat{x}^{+}_{\phi})=\bigl(D_{\phi}\mu_{\phi}(x)\bigr)^{\top}\Delta\hat{G}(x)^{\top}\nabla_{x}B_{\theta}(\hat{x}^{+}_{\phi})

Reviewer_jytMOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces NBF-PAS, a framework for learning neural control barrier functions without relying on an external safe-control oracle. The proposed approach uses low-dimensional safety projections and recovery searches along failed trajectories to identify anchor states that guide barrier learning. These anchors provide supervision for both the location and orientation of the learned safety boundary. The framework is evaluated on Dubins, Segway, quadrotor, and Hopper systems, demonstrating improvements in collision avoidance and, following policy retraining, task performance compared to the considered baselines.

Critical strengths

The paper addresses a relevant challenge in neural barrier learning, namely the difficulty of identifying appropriate safety boundaries without an external safe-control oracle. The use of recovery trajectories to guide boundary placement and orientation is an interesting approach that leverages information directly available from system interactions.

The experimental evaluation covers systems with different dynamics and state dimensions, including relatively complex robotic systems. The improvements in collision avoidance are encouraging, particularly the Hopper results, where barrier-guided policy retraining substantially improves goal-reaching performance compared to the pure RL baseline.

Critical weaknesses

The connection between recovery anchors and the actual controlled-invariant boundary requires further justification. The recovery criterion in Equation (16) is based on the alignment of projected state displacement with the obstacle normal. However, satisfying this criterion does not necessarily imply that a state is recoverable over a longer horizon or lies near the boundary of the viability kernel. In particular, the negative acceptance thresholds reported in Table 4 allow transitions that initially move toward the obstacle to be accepted. It is therefore unclear how reliably these anchors identify the desired safety boundary.

The formulation and safety guarantees remain incomplete. The proposed loss uses a projected barrier gradient, although Appendix E acknowledges that this quantity is not uniquely defined for general non-invertible projections. This discrepancy should be resolved, particularly for systems such as Hopper. Furthermore, the theoretical discussion establishes conditional forward-invariance requirements rather than guarantees for the learned barriers. The experiments report collision rates but do not directly evaluate whether the learned CBF conditions hold over the relevant state domains. This limits the strength of the safety claims.

Questions for authors

How does the proposed recovery criterion ensure that the identified anchors approximate the controlled-invariant boundary? In particular, how are successful recovery trajectories distinguished from transitions that merely satisfy the projected directional threshold?

How is the projected gradient used in the anchor loss computed for non-invertible safety projections? Appendix E appears to acknowledge that this quantity is not generally well-defined, which seems inconsistent with the main training formulation.

Can the authors provide an evaluation of CBF constraint violations on held-out states, together with an ablation isolating the contributions of anchor placement and normal alignment? This would help establish whether the improvements arise from better barrier geometry rather than additional recovery-based supervision alone.

4 comments

Weakness 1-1:

Thank you for the point. We agree that our proposed anchors construction does not imply long-horizon safety invariance. However, these anchors remain useful because they provide soft geometric targets around which the barrier’s invariance conditions are fitted.

Each anchor specifies a preferred boundary location through placement penalty min⁡θBθ(xt+1)2\min_{\theta} B_{\theta}(\mathbf{x}_{t+1})^2, not a state that the learned safe set must include, while normal alignment min⁡θ(n^(xt+1)⊤nθ(xt+1))2\min_{\theta} (\hat n(\mathbf{x}_{t+1})^\top n_{\theta}(\mathbf{x}_{t+1}))^2 supplies a preferred local orientation. Importantly, both are soft penalties optimized jointly with the full-state CBF conditions, so training seeks a boundary that reconciles the anchor’s geometric guidance with the dynamical conditions for invariance. Recovery anchors thus gives the learner explicit geometric targets beyond the unsafe-state labels supplied by terminal failures.

We added new ablation in Appendix D.2 to examine the contribution of this anchor supervision. Removing both placement and alignment disables the use of anchor states, while retaining the standard neural CBF fitting. The resulting model assigns negative barrier values throughout the displaced (xy) slice, whereas the full objective retains a localized, closed boundary. This illustrates the value of anchor supervision: it helps preserve boundary structure that minimizing the standard empirical losses through training demonstrations alone does not ensure. The table below shows the collision rate, barrier violation rate, and unsafe area across 5 seeds in Quadrotor system. Specifically, barrier violation rate computes the rate of states that fail the CBF condition among visited states, and unsafe area is the fraction of state-space grid area where B(x)<0B(x)<0 for all visited states.

SettingsCollision RateBarrier Violation RateUnsafe Area
NBF-PAS1.9% ± 1.113.2% ± 3.226.6% ± 7.0
No Anchors14.8% ± 3.39.8% ± 3.944.2% ± 18.3
No Normal Alignments6.4% ± 4.411.9% ± 4.835.0% ± 6.2
No Both26.9% ± 1.013.2% ± 2.668.9% ± 3.6
Baseline27.1% ± 8.026.7% ± 4.966.3% ± 17.6

Recall that the baseline minimizes the ∇Bθ⊤x˙+γBθ\nabla B_{\theta}^{\top}\dot{\mathbf{x}}+\gamma B_{\theta} while no both only minimizes ∇Bθ⊤x˙\nabla{B}_\theta^\top \dot{\mathbf{x}}.

Weakness 1-2 (negative angle)

We allow a negative threshold because even trajectories on a true invariant safety boundary can move toward an obstacle in workspace. For example, in Appendix F, we show a state on the numerical Hamilton-Jacobi Reachability zero contour whose velocity points partly toward the obstacle (n^⊤v^=−0.607\hat n^\top\hat v=-0.607). Negative projected alignment therefore does not imply crossing the safety boundary: invariance depends on the joint evolution of position and velocity. Requiring non-negative alignment would exclude such valid boundary transitions. Allowing ρ<0\rho < 0 keeps these candidates eligible as soft geometric targets, while the full CBF conditions remain unchanged.

Weakness 2:

Thanks for identifying the inconsistency in Appendix E. We have corrected this passage (revised in red) to match the formulation of the main text and implementation. We train our barrier functions in full state and do not reconstruct states from the projected coordinates. Hence, the non-invertibility does not affect the barrier learning process, while the projected anchors can still guide the search in the full state space.

To define this one-way trajectory precisely, we use the Jacobian-based mapping to compute projection (Appendix E). Let Jκ(x)=Dκ(x)J_\kappa(x)=D\kappa(x). We use the workspace alignment vector qλ(x)=(Jκ(x)Jκ(x)⊤+λI)−1Jκ(x)∇xBθ(x)q_\lambda(x) = \bigl(J_\kappa(x)J_\kappa(x)^\top+\lambda I\bigr)^{-1}J_\kappa(x)\nabla_x B_\theta(x) that solves the regularized least-squares problem \min_{q}\\;\left\\|J_{\kappa}(x)^{\top}q-\nabla_{x}B_{\theta}(x)\right\\|_{2}^{2}+\lambda\\|q\\|_{2}^{2}. That is, Jκ(x)⊤qJ_\kappa(x)^\top q is the approximation of full-state barrier gradient. We use Jacobian because it captures how multiple configuration variables jointly move the safety-relevant body part.

Question 1:

As discussed in our response to weakness 1, our anchors construction does not imply long-horizon safety invariance. However, these anchors remain useful because they provide soft geometric targets around which the barrier’s invariance conditions are fitted. The recovery search and backtracking provide selected full states with dynamically realizable transitions (xt,ut,xt+1,κ(xt+1)−κ(xt))∣(κ(xt+1)−κ(xt))⊤n^(xt+1)≥ρ\\{(\mathbf{x}_{t},\mathbf{u}_{t},\mathbf{x}_{t+1},\kappa(\mathbf{x}_{t+1})-\kappa(\mathbf{x}_{t}))\mid(\kappa(\mathbf{x}_{t+1})-\kappa(\mathbf{x}_{t}))^{\top}\hat{n}(\mathbf{x}_{t+1})\geq\rho\\}: states guide where to place a candidate zero level set (min⁡θBθ(xt+1)2\min_{\theta}B_{\theta}(\mathbf{x}_{t+1})^2), and transitions guide how to orient it locally (min⁡θ(n^(xt+1)⊤nθ(xt+1))2\min_{\theta}(\hat{n}(\mathbf{x}_{t+1})^{\top}n_{\theta}(\mathbf{x}_{t+1}))^2). This is the key recipe that gives control-dependent supervision for shaping the barrier beyond classification.

Question 2:

We do not invert low-dimensional projections to full states, so non-invertibility is not a bottleneck in our approach. As discussed in weakness 2, we use Jacobian-based mapping to compute projection so that Jκ(x)⊤qJ_\kappa(x)^\top q approximates the full state barrier gradient. Thus, we obtain an alignment vector even when the projection is non-invertible. It does not assume that BθB_\theta is a function of projected coordinates alone.

Question 3:

We added an ablation test isolating the contribution of anchor placement and normal alignment with the Quadrotor example on 5 seeds in Appendix D.2. We include the conventional training that minimizes ∇Bθ⊤x˙+γBθ\nabla B_{\theta}^{\top}\dot{\mathbf{x}}+\gamma B_{\theta} as baseline. Removing anchor placement increases the collision rate from 1.9% to 14.8% and the fraction of the evaluated region labeled unsafe from 26.6% to 44.2%. Removing normal alignment increases collisions to 6.4% and the unsafe-labeled fraction to 35.0%. Removing both produces 26.9% collisions and labels 68.9% of the evaluated region unsafe—approximately 2.6 times the fraction for NBF-PAS. The accompanying visualizations show that the full objective preserves a localized boundary, whereas removing these objectives produces substantially broader negative-valued regions.

We thank the authors for their detailed response and additional ablation experiments. The explanation for allowing negative alignment thresholds is reasonable, and the new results provide useful evidence that anchor supervision improves empirical collision avoidance. However, our concerns regarding the mathematical formulation and safety guarantees remain. The Jacobian-based least-squares construction provides a projected alignment surrogate, but does not generally establish a uniquely defined projected barrier gradient for non-invertible projections. Furthermore, the proposed method still exhibits a 13.2% CBF violation rate on visited states, compared with 9.8% without anchors, despite achieving lower collision rates. Can the authors clarify the precise mathematical definition of the projected gradient used in training and explain this discrepancy between empirical collision avoidance and satisfaction of the CBF conditions?

Projection:

Thank you for the follow-up. Our construction does not require an inverse projection -- only forward projection from full states x\mathbf{x} to the low-dimensional κ(x)\kappa(\mathbf{x}) is used and this mapping is deterministic. The alignment of κ(x)\kappa(\mathbf{x}) with the normal direction n^(xt+1)=qλ(xt+1)∥qλ(xt+1)∥2\hat{n}(\mathbf{x}_{t+1})=\frac{q_\lambda(\mathbf{x}_{t+1})}{\lVert q_{\lambda}(\mathbf{x}_{t+1})\rVert_2} is uniquely defined by their dot product.

We understand that the reviewers' concern is that this alignment does not uniquely define the zero-level set of the barrier function, which is true, and it is our intended design. The motivation of our approach is to avoid the use of any additional supervision in the training of neural barriers, and this low-dimensional alignment is what we believe to be a minimal and natural learning signal for bootstrapping the neural barrier training from the intended safety properties themselves. With this singled-value mapping, we can turn observable body-point motion into explicit geometric supervision for the full-state barrier.

As we compared in the paper, without using this low-dimensional alignment signal, standard training fits the neural barrier directly with full state constraint, requiring massive datasets without geometric guidance on how nearby states should orient the barrier. Our proposed NBF-PAS prevents this blind search in the full-state. For instance, in Hopper, different joint configurations and velocities can produce similar body-point motion relative to the obstacle. NBF-PAS can identify these transition that share a geometric supervision rule without assigning them the same barrier value. This gives the network an explicit direction for shaping the boundary at each anchor instead of only relying on sampled CBF inequalities for orientation inference.

Violation rate and collision avoidance:

This discrepancy actually highlights that the convergence on the standard barrier condition objective can mislead the learning of barriers, in the sense that low violation of the training objective does not imply low collision in the trained barrier, which is a core difficulty of existing training methods and the motivation of our use of projected anchors.

Note that the violation rate alone does not measure whether the safe region is useful for control. The evaluation checks local feasibility at on-boundary states visited by each model's own trajectories. No-anchor obtains a lower rate because it learns a more conservative barrier: it assigns a configuration a sufficiently negative barrier value that moves this hard configuration outside evaluation set. In our ablation test, removing anchor placement expands the negative-barrier area from 26.6% to 44.2%, alongside an increase in collisions from 1.9% to 14.8%. Hence, a low sampled violation rate can therefore give a misleading interpretation of the learning process when the barrier excludes useful configurations and provides poor control guidance, while our NBF-PAS enforces additional geometric supervision through direct anchor placement and orientation to learn a more useful barrier for control.

Comments on the paper

1 comment

We thank the reviewers for their constructive feedback. We added the following ablations and explanations in our paper:

  • Appendix D.2: Five-seed Quadrotor ablations isolating anchor placement and normal alignment.
  • Appendix D.3: PNCBF comparisons using an unseen MPC controller on 400 held-out initial-state–goal pairs.
  • Appendix D.4: Four-seed CPO comparisons with scratch and nominal initialization, including variance reporting.
  • Appendix D.5: Two-obstacle Segway experiments reusing the frozen single-obstacle barrier.
  • Appendices E–F: A corrected definition of the Jacobian-based alignment vector and a numerical Dubins HJR example explaining negative radial alignment.

These additions highlight the benefit of our anchor supervision: it guides both where to place a candidate boundary and how to orient it, helping learn a useful barrier landscape beyond safe/unsafe classification.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.