Control Barrier Functions for Safety-Aware Latent-Space Motion Planning

Chandrakiran Ravula, Siddhanth RamesharXiv 2026 · arXiv

TL;DR

This paper introduces a novel approach to robotic motion planning by replacing standard binary collision classifiers in latent-space planners with Neural Control Barrier Functions (CBFs). By enforcing safety through a closed-form "predictor-corrector" step, the authors achieve significantly higher success rates and smoother trajectories for high-degree-of-freedom manipulators compared to traditional gradient-based methods.

What Problem Does This Paper Address?

Motion planning for robots like the 7-DoF Franka Panda is computationally expensive due to high dimensionality and complex kinematic constraints. Latent-Space Path Planning (LSPP) offers a solution by using Variational Autoencoders (VAEs) to compress the robot's configuration space into a low-dimensional, differentiable manifold.

However, existing LSPP frameworks rely on binary collision classifiers to avoid obstacles. These classifiers suffer from two major flaws:

  1. Poor Gradient Landscape: Their sigmoidal outputs create "cliff-edge" gradients—zero far from obstacles and spiking sharply near them—which often trap optimizers in local minima.
  2. Lack of Temporal Structure: They evaluate static states but don't guarantee that a safe state at step $k$ won't transition into a collision at step $k+1$.

Methods

The authors propose a Latent-Space Control Barrier Function to replace the classifier.

1. The Architecture

The system uses a VAE to map the robot's state (joint angles and end-effector position) into a 7D latent space. A neural network is then trained to act as a CBF ($B(z)$) directly within this latent space. Unlike a classifier, the CBF is an unbounded, signed function where $B(z) \geq 0$ defines the "safe set."

Fig. 2. Baseline LSPP architecture. Top: The VAE compresses the robot state (q ∈R7, e ∈R3) into a 7-D latent space regularized by a prior. Bottom: The binary collision classifier predicts collision from the latent code concatenated with the obstacle descriptor o ∈R4.

2. Training the Barrier

The CBF is trained using a three-term objective:

  • Safe/Unsafe Sign Constraints: Ensuring the function is positive in free space and negative in collision.
  • Discrete Decrease Condition: Learning how the safety margin changes over time based on the transitions of an "obstacle-blind" nominal planner.

3. Predictor-Corrector Safety Filter

At planning time, the authors decouple goal-seeking from safety. The planner first takes a "nominal" step toward the goal. If this step violates the CBF condition, a closed-form, first-order correction projects the latent state back onto the safety constraint along the barrier gradient. This avoids the "gradient interference" common when safety and goal-reaching are summed into a single loss function.

Datasets

The models were trained and evaluated using:

  • State-label dataset: 100,000 free-space transitions and 100,000 collision states for a Franka Panda robot.
  • Transition dataset: Pairs of states produced by an obstacle-blind planner to teach the CBF the system's dynamics.
  • Evaluation Scenarios: Three settings of increasing complexity: Fixed obstacle (S1), Multi-position obstacles (S2), and General random obstacles (S3).

Key Insights and Ideas

  • Gradient Quality Matters: The CBF provides dense, non-vanishing gradients that guide the optimizer even when far from an obstacle, preventing the "stalling" seen with binary classifiers.
  • The Power of Modularity: By treating safety as a filter (projection) rather than a penalty term in the loss, the planner can pursue the goal aggressively without safety terms pulling it into local minima.
  • Implicit Regularization: Surprisingly, the model performed better on the multi-position task (83.8% success) than the fixed-obstacle task (75.6%). The authors suggest that learning to generalize across different obstacle positions forces the network to learn a smoother, more robust barrier surface.
  • The $\alpha$ Knob: The decay rate $\alpha$ acts as a single, interpretable hyperparameter to tune the safety-performance trade-off. Low $\alpha$ is overly conservative; high $\alpha$ allows risky steps that break the first-order approximation.
Fig. 1. Top: A naive trajectory executed without collision avoidance intersects the obstacle. Bottom: A safe trajectory guides the manipulator over the obstacle to the target.

Key Results

  • Success Rates: On a fixed-obstacle task, the CBF reached a 75.6% success rate, outperforming the best-tuned classifier by 12 percentage points.
  • Efficiency: The CBF planner averaged 94.8 ms per query, roughly twice as fast as the reproduced baseline (186.3 ms), largely because it avoids the optimization stalls caused by poor gradients.
  • Safety-Performance Trade-off: The best results were consistently found with $\alpha$ between 0.10 and 0.20.
Table showing success rates across different alpha values

Limitations and Future Work

The primary limitation identified is covariate shift. In the fully general random-obstacle setting (S3), success dropped to 33%. The CBF, trained on "blind" transitions, struggled when encountering the specific states visited by the corrected planner at deployment.

Future directions include:

  • Iterative Data Aggregation: Using techniques like DAgger to retrain the barrier on states visited by the safety-aware planner.
  • Latent MPC: Moving from open-loop trajectories to a closed-loop Model Predictive Control formulation to better handle VAE decoding errors.
  • Generative Safety: Integrating CBF constraints into diffusion-based planners.

Summary updated

Abstract

Latent-space path planning (LSPP) compresses the configuration space of a robotic manipulator into a compact, differentiable manifold learned by a variational autoencoder, enabling fast gradient-based planning without explicit inverse kinematics. Existing LSPP frameworks enforce obstacle avoidance with a binary collision classifier, whose saturated gradients near obstacle boundaries and lack of temporal structure cause the optimizer to stall and provide no forward-invariance guarantee. We replace the classifier with a neural control barrier function (CBF) learned directly in the latent space. The barrier is trained with a three-term objective that combines safe and unsafe sign constraints with a discrete-time decrease condition evaluated on transitions of an obstacle-blind nominal planner, and it is enforced at planning time by a closed-form, first-order correction that projects each goal-directed latent step back onto the CBF constraint. On a 7-DoF Franka Panda manipulator, the CBF planner achieves a 75.6\% success rate on a fixed-obstacle task, 12.0 percentage points above the best of 35 tuned classifier configurations, and 83.8\% on a multi-position task. A sweep of the decay rate $α$ characterizes the safety--performance trade-off, with the best operating points in $α\in [0.10, 0.20]$. In a fully general random-obstacle setting, success drops to 33.0\%, which we trace to covariate shift between the obstacle-blind training transitions and the states visited by the corrected planner.
BibTeX
@misc{ravula2026control,
  author = {Chandrakiran Ravula and Siddhanth Ramesh},
  title = {{Control Barrier Functions for Safety-Aware Latent-Space Motion Planning}},
  year = {2026},
  eprint = {2611.01814},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.01814},
}
Figure