Skip to main content
Back to timeline
arXivSource publication:

Sim2CBF drops model derivatives from the safety filter: forward simulation alone keeps a box-toppling robot safe

Synopsis

Robot failure avoidance often relies on control barrier functions, but solving the safety filter usually requires derivatives of the predictive model, which contact-rich simulators and learned world models often cannot supply reliably. Sim2CBF approximates the barrier constraint with finite differences and local linear interpolation over sampled control inputs, turning the original optimization into a quadratic program that needs only forward evaluations while retaining failure-avoidance guarantees from a backup policy. In simulated box-toppling it reaches a high safety rate across trials; on a latent-space Dubins car task its safe success rate is 0.95, above direct blending at 0.86 and sample-and-select at 0.82, but below a neural barrier-function baseline at 1.00.

Source-provided article image: Sim2CBF: From Forward Simulations to Implicit Control Barrier Function Safety Filtering without Model Derivatives
arXiv

Interpretation

The safety filter can be solved without computing derivatives of the predictive model, correcting a nominal control away from failure states using forward simulation alone. Earlier approaches either solved the safety filter with gradient methods, requiring the Jacobian of the dynamics with respect to the control input, or fell back to pure sample-and-select, whose performance depends on the sampling strategy and degrades with control dimension. On a 1D double integrator, analytical dynamics, MJX automatic differentiation and MJWarp forward evaluation were compared for computation time; automatic differentiation became substantially slower when contact dynamics were simulated, while MJWarp had lower computation time when modeling contacts. On a 2D double integrator collision-avoidance task, analytical, sample-and-select and interpolation with blending all avoided the obstacle and progressed toward the goal.

The method replaces the barrier-constraint derivative with finite differences and interpolates linearly over the convex hull of sampled control inputs, yielding a convex quadratic program. The finite-difference constraint is nonlinear in the control input and still needs derivatives to solve directly; interpolation makes the constraint linear in interpolation weights, so only sampled function values are used. Candidate controls include the nominal input and a backup input that satisfies the original constraint by construction, augmented with small task-space perturbations; on a simulated 7-DoF Franka Panda box-toppling task, interpolation with blending reached a high safety rate over trials.

The implicit barrier function is built from finite-horizon rollouts of a backup policy, so changing the predictive model, input constraints or failure set needs no retraining. Neural barrier functions parameterize the safe set in a network and must be retrained when the model or failure specification changes; the implicit construction leaves safety to rollout evaluation. On the latent-space Dubins car task, interpolation with blending reached a safe success rate of 0.95, direct blending 0.86, sample-and-select 0.82 and neural CBF with sample-and-select 1.00; under the single-observation setting, interpolation with blending stayed at 0.94 while the neural CBF approach fell to 0.11.

Perspective

The demonstrated scope covers a 1D double integrator, simulated box-toppling with a 7-DoF Franka Panda, hardware box-toppling with a UFACTORY xArm 7, and Dubins car collision avoidance in the latent space of a learned world model; in the hardware experiments the box dimensions and weight are known and its pose is estimated with AprilTags.

The finite-difference condition guarantees failure avoidance only at prediction instants, not between consecutive samples; the interpolated constraint approximates the original nonlinear constraint only within the convex hull of sampled controls, and the error typically grows with hull size. In the hardware experiments, perception errors make the constructed barrier function conservative, and the authors note that improved perception would reduce this conservatism.