As robots are increasingly deployed in everyday environments, ensuring their safety has become a central challenge. Existing methods often encode safety requirements as opaque mathematical/logical formulations or dense cost functions. While effective in specific tasks, they remain difficult to interpret, tightly coupled to individual tasks, and offer limited insight into why a robot action is considered safe or unsafe. To address this limitation, we propose ``Neuro-Symbolic Predicate Learning for Semantic Safe Robot Control'' (NEUPRO), which leverages a differentiable reasoner that can learn reusable safety representations from human-specified safety knowledge. NEUPRO allows practitioners to express task-related safety requirements as transparent symbolic rules, while enabling gradients to propagate through these rules to a feature extractor that maps raw observations to safety-relevant concepts. As a result, the learned feature extractor is (softly) grounded in human-understandable semantics, supports transparent constraint evaluation, and is transferable across tasks. By coupling interpretability with differentiability, NEUPRO moves beyond opaque cost design toward reusable safety reasoning. To evaluate NEUPRO's capability, we collect and release REASON, the first real robot benchmark dataset for interpretable robot safety specification. Experiments on REASON show that NEUPRO learns safety-critical features that generalize across tasks, mitigate the interpretability limitations of conventional black-box cost formulations, and provide explicit explanations of safety violation.
Figures & tables
Figure 1: Existing safe robot learning methods often encode safety constraints using opaque mathematical/logical functions, which are hard to interpret and bound to specific tasks. We propose NEUPRO, leveraging differentiable reasoning to alleviate these limitations using interpretable symbolic rules.
Method
Interp.
Veri.
Flexi.
Math. constraints
✗
✓
✗
Learning-based
✗
✗
⚫
VLM-based
⚫
✗
✓
NEUPRO
✓
✓
✓
Table 1: NEUPRO alleviates the limitations of existing safety reasoning paradigms by supporting interpretable, verifiable, and flexible safety reasoning. We compare different paradigms along these three dimensions, where ✓ , ⚫ , and ✗ indicate strong, partial, and limited support, respectively.
Figure 2: Overview of NEUPRO . NEUPRO supports interpretable safety reasoning, end-to-end grounding of safety concepts in images, and flexible reasoning with commonsense knowledge. NEUPRO consists of two key components: (i) a neural perception module and (ii) a graph-based differentiable reasoner. Given a raw image, the perception module uses Grounding DINO and a learnable MLP to extract object-centric features and predict soft truth values for safety-relevant grounded atoms. The differentiable reasoner then infers the safety prediction based on the ground atom evaluation, safety rules, and commonsense knowledge. For details, please see Sec. 3 .
Figure 3: NEUPRO accurately learns predicates through differentiable reasoner supervision from raw images. NEUPRO outperforms baseline methods in predicate grounding accuracy. For readability, only mean accuracy is shown. Details see RQ1.
Method
Tasks
Occlusion
Cabinet
Near
Above
Collision
Close
Pointing
All task
Black-box classifier
0.84 ± 0.09
0.96 ± 0.04
0.79 ± 0.08
0.98 ± 0.05
0.96 ± 0.05
0.98 ± 0.04
0.86 ± 0.14
N/A
DeepSeek-VL2 (ICL)
0.41 ± 0.11
0.5 ± 0.11
0.54 ± 0.07
0.3 ± 0.24
0.31 ± 0.07
0.4 ± 0.06
0.33 ± 0.16
0.48 ± 0.03
Qwen3.5 (ICL)
0.49 ± 0.02
0.61 ± 0.02
0.39 ± 0.02
0.68 ± 0.16
0.52 ± 0.14
0.82 ± 0.04
0.57 ± 0.16
0.55 ± 0.04
DeepSeek-VL2
0.41 ± 0.01
0.51 ± 0.06
0.53 ± 0.04
0.31 ± 0.09
0.39 ± 0.02
0.21 ± 0.06
0.31 ± 0.03
0.44 ± 0.04
Qwen3.5
0.5 ± 0.02
0.62 ± 0.01
0.39 ± 0.0
0.78 ± 0.03
0.75 ± 0.0
0.8 ± 0.11
0.77 ± 0.02
0.55 ± 0.04
Table 2: NEUPRO accurately infer safety outcomes from raw visual observation. Safety classification accuracy comparison across all REASON tasks with baseline methods (the higher, the better, best performing bolded). Results averaged over five test groups with std, details see RQ2. ICL denotes in-context learning.
Figure 4: Qualitative example of safety violation explanations produced by NEUPRO , in the form of first-order logic rules and atoms. In this scene, the grounded atoms indicate that the robot is holding scissors and is close to a person, from which NEUPRO infers an unsafe prediction through the corresponding first-order logic rule.
Figure 5: NEUPRO learned predicates are transferrable. Object-level generalization evaluates whether predicates remain accurate when applied to different object instances or categories. Task-level generalization compares predicates trained directly on a target task with predicates transferred from another task. NEUPRO retains high accuracy under both settings, demonstrating that its learned representations are reusable across objects and tasks. Details see RQ4.
Method
Above_Laptop
Close_Person
Hard
Soft
Sharp
N.sha
NEUPRO (ours)
1.0 ± 0.0
1.0 ± 0.0
1.0 ± 0.0
1.0 ± 0.0
Qwen3.5
0 ± 0.0
0.96 ± 0.08
0.04 ± 0.08
0.96 ± 0.08
Qwen3.5(ICL)
0.36 ± 0.15
0.76 ± 0.08
0.6 ± 0.28
0.92 ± 0.10
DeepSeek-VL2
0.64 ± 0.15
0.32 ± 0.20
0.72 ± 0.16
0.36 ± 0.08
DeepSeek-VL2(ICL)
0.56 ± 0.29
0.44 ± 0.32
0.84 ± 0.08
0.4 ± 0.22
Table 3: NEUPRO supports flexible safety reasoning with commonsense knowledge. We evaluate whether each method distinguishes identical spatial relations with different object properties: hard vs. soft objects above a laptop, and sharp vs. non-sharp objects close to a person. Results are reported as mean accuracy with std. See RQ5 for details.
Figure 6: NEUPRO supports scalable training and inference through graph-based reasoning. We compare graph-based NEUPRO with its tensor-based variant. NEUPRO achieves 9.1× faster training, 14.3× lower memory usage and 1.6× faster inference, demonstrating substantially better scalability. Details see RQ6.
Physical interactions can create future hazards that are not apparent from the robot's current geometric surroundings. We present a framework termed Predictive Semantic Safety (PSS), which connects visual physical reasoning to backup-based safety filtering. A vision-language model (VLM) predicts physical events and their timing or directly predicts object displacements. An explicit motion model converts event hypotheses into object trajectories. Split conformal prediction calibrates position errors jointly across specified objects, observation times, and future times; geometric shape bounds convert the resulting position regions into predicted object occupancy. PSS evaluates a prescribed backup maneuver against this occupancy and derives input-affine constraints for minimally modifying the nominal input while preserving backup feasibility under the robot dynamics and input limits. MuJoCo experiments with a Unitree Go1 consider falling fixtures, impact-driven support loss, and contact propagation. PSS achieves a safe episode rate of 99.3%, compared with 43.3% for a Backup Control Barrier Function baseline that only uses current obstacle geometry.
Taekyung Kim, Salem Fradi, Yanning Dai +2
Department of Robotics, University of Michigan, Ann Arbor, MI 48109, USA · Center of Excellence in Generative AI, King Abdullah University of Science and Technology (KAUST), Thuwal, Saudi Arabia. · Dalle Molle Institute for Artificial Intelligence Research (IDSIA), Switzerland. +2
Deploying foundation models for robot control raises a central challenge: the expressive power that enables rich, multimodal perception also makes these models opaque and difficult to analyze formally, rendering them intractable for existing verification tools. In this paper, we present FEARL (Foundation-Enabled Assured Robot Learning), a framework that addresses this tension through a modular architectural decomposition. FEARL separates the policy into a large Controller (C) responsible for high-dimensional perception and task reasoning, and a small Safety module (S) that receives low-dimensional observations from dedicated safety sensors together with a bounded context embedding from C and produces the final action. Since many robot safety requirements, such as collision avoidance and workspace boundary constraints, can be expressed over these safety sensor observations, formal verification can be applied to S rather than to the full foundation-model backbone. This makes formal analysis tractable with existing tools while preserving the Controller's expressive power for task reasoning. To show that the decomposed policy remains capable of solving diverse tasks, we evaluate FEARL on three simulated robotic domains using multiple Controller backbones and training procedures, including pretrained off-the-shelf vision-language-action models. We further transfer the learned policy from one of our simulated tasks to a physical robot, suggesting that the low-dimensional safety interface supports practical sim-to-real transfer.
Traditional safety-critical control methods, such as control barrier functions, suffer from semantic blindness, exhibiting the same behavior around obstacles regardless of contextual significance. This limitation leads to the uniform treatment of all obstacles, despite their differing semantic meanings. We present Safe-SAGE (Social-Semantic Adaptive Guidance for Safe Engagement), a unified framework that bridges the gap between high-level semantic understanding and low-level safety-critical control through a Poisson safety function (PSF) modulated using a Laplace guidance field. Our approach perceives the environment by fusing multi-sensor point clouds with vision-based instance segmentation and persistent object tracking to maintain up-to-date semantics beyond the camera's field of view. A multi-layer safety filter is then used to modulate system inputs to achieve safe navigation using this semantic understanding of the environment. This safety filter consists of both a model predictive control layer and a control barrier function layer. Both layers utilize the PSF and flux modulation of the guidance field to introduce varying levels of conservatism and multi-agent passing norms for different obstacles in the environment. Our framework enables legged robots to safely navigate semantically rich, dynamic environments with context-dependent safety margins.