Control Barrier Functions Are Having a Moment, But the Hard Problems Remain Unsolved

This is genuinely new in the sense that prior work on CBFs for humanoids has largely ignored the kinematic-to-dynamic transfer problem, or assumed it away. The experiments are also more comprehensive than typical: locomotion, teleoperation, and single-leg balancing with hand control. The project website shows videos of a real humanoid maintaining safety constraints during these tasks, which is more than many papers in this space deliver.

But here is where I have to note some limitations. The "bounded disturbances" assumption is doing a lot of work. How bounded? The paper does not provide quantitative characterization of what happens when disturbances exceed the assumed bounds. The real-robot experiments are impressive but the sample size is small, and it remains unclear how the conservative tuning affects task performance in scenarios where safety margins are tight. These are not fatal flaws, but they are the kind of details that matter when you are trying to assess whether this approach will generalize beyond the lab.

Learning dynamics from data sounds great until you try it. The second paper takes a different approach to the same fundamental problem (arXiv). Instead of assuming known dynamics and tuning conservatively, the researchers propose learning a Koopman predictor from data and constructing CBF constraints in the lifted Koopman space. For those unfamiliar with Koopman operators, the basic idea is that any nonlinear dynamical system can be represented as a linear system in a (possibly infinite-dimensional) space of observables. If you can find a good finite-dimensional approximation, you get the benefits of linear control theory for nonlinear systems.

The framework, which they call Robust Koopman-CBF SAC, combines this learned dynamics model with soft actor-critic reinforcement learning. The safety filter is a quadratic program that enforces the CBF constraints, and crucially, they account for approximation error by tightening the CBF condition using a projected residual margin estimated from held-out data. The actor is regularized toward the feasible set, which is a sensible way to reduce dependence on the safety filter over training.

It's worth noting that this addresses a real limitation of existing safe RL methods: most either require accurate dynamics models (which you rarely have) or hand-designed barrier certificates (which require expertise and don't generalize). Learning both the dynamics and the safety constraints from data is appealing, at least in principle.

The results are mixed in ways that are actually informative. On CartPole stabilization and tracking, the method achieves zero constraint violations while matching unconstrained SAC returns. That's a clean result on a simple benchmark. On the Safety Gymnasium locomotion tasks, which are higher-dimensional and more complex, the picture is messier. The method reduces violations in some settings but, and I appreciate that the authors are honest about this, "exposes important limitations of first-order velocity barriers and linear EDMD models."

This is the kind of negative result that rarely gets highlighted but matters enormously. First-order velocity barriers, which constrain how fast the system can approach unsafe states, are a common choice because they are easy to implement. But they can be overly conservative or, worse, insufficient for systems with significant inertia. Linear Extended Dynamic Mode Decomposition (EDMD) models can capture some nonlinear dynamics but struggle with highly nonlinear systems. The authors note that these limitations motivate "high-order and multi-step Koopman-CBF extensions," which is research-speak for "we haven't solved this yet."

The gap between benchmarks and reality. Reading these papers together highlights a pattern that runs through much of robot safety research. The mathematical frameworks are increasingly sophisticated. The benchmark results are increasingly impressive. And yet the gap between what we can prove in simulation and what we can guarantee on a real robot doing real tasks remains enormous.

Consider what "safety" actually means for a humanoid robot operating in a human environment. It is not just about joint limits and self-collision. It is about not knocking over a child who runs into the room unexpectedly. It is about recognizing when a surface is unstable before stepping on it. It is about understanding that the box you are carrying contains fragile items. None of this is captured by the safety constraints in these papers, nor could it be, because we do not have good formal specifications for these higher-level safety requirements.

This is not a criticism of the research itself. Both papers make genuine contributions to the specific problems they address. The ISSf-CBF framework provides a principled way to handle the kinematic-to-dynamic transfer problem. The Koopman-CBF approach offers a path toward learning-based safety that does not require hand-designed barrier functions. These are incremental advances over prior work, but incremental advances are how fields make progress.

The problem is how this research gets interpreted. It is too early to say whether CBF-based approaches will scale to the full complexity of real-world robot safety. The experiments in these papers involve controlled environments, known obstacle locations, and relatively simple task specifications. Extending to unstructured environments with dynamic obstacles and uncertain task requirements is, well, it's basically the entire remaining problem.

What I'd want to see next. If I were reviewing follow-up work in this area, here is what I would be looking for.

First, quantitative characterization of failure modes. Both papers demonstrate that their methods work in the tested scenarios. Neither provides systematic analysis of when and how they fail. What happens when disturbances exceed the assumed bounds? What happens when the Koopman approximation error is larger than expected? Understanding failure modes is arguably more important than demonstrating success cases.

Second, comparison across methods. The field would benefit from standardized benchmarks that allow direct comparison between different safety approaches. The Safety Gymnasium tasks are a start, but they are not comprehensive, and different papers use different metrics. It remains unclear whether ISSf-CBFs would outperform Koopman-CBFs on the same tasks, or vice versa, because no one has done the comparison.

Third, integration with perception. Both papers assume that the robot knows the locations of obstacles and can measure its own state accurately. This is reasonable for initial work, but perception uncertainty is a major source of safety failures in real systems. How do these frameworks degrade when state estimation is noisy or obstacle detection has false negatives?

Fourth, and this is perhaps the hardest, formal treatment of the specification problem. CBFs guarantee that the system stays within some safe set, but who defines that set? How do we know the specification is complete? The papers assume that safety constraints are given, but in practice, writing down the right constraints is often harder than enforcing them.

The honest assessment. Control barrier functions are a promising approach to robot safety. The two papers discussed here represent solid incremental progress. The ISSf-CBF framework handles disturbances more robustly than standard CBFs. The Koopman-CBF approach reduces dependence on hand-designed models. Both include real experiments or realistic simulations that demonstrate the methods work in at least some scenarios.

But we should be clear-eyed about what has not been solved. Safety guarantees that hold under bounded disturbances are not the same as safety guarantees that hold in the real world, where disturbances are often unbounded or at least poorly characterized. Learning dynamics from data is useful, but learned models can fail in unexpected ways, and the safety implications of model failure are not well understood. Benchmark success does not imply real-world success.

The researchers themselves seem to understand this. The Koopman-CBF paper explicitly notes the limitations of their approach and suggests directions for future work. The ISSf-CBF paper is careful to describe their guarantees as conditional on the assumed disturbance bounds. This kind of intellectual honesty is valuable, even if it does not make for exciting headlines.

Actually, the research shows something important that often gets lost in coverage of robot safety work: we are still in the early stages of understanding how to make robots safe in any rigorous sense. The mathematical tools are improving. The experimental validation is becoming more thorough. But the fundamental problem, keeping robots from hurting people in complex, unstructured environments, remains largely unsolved. These papers are contributions to that larger project, not solutions to it.

For now, the honest answer to "are robots safe?" is "it depends on what you mean by safe, and in what context, and under what assumptions." That is not a satisfying answer, but it is the accurate one. The work continues.

Aisha Patel · 4 hours ago · 9 min