SafeManip: A Property-Driven Benchmark for Temporal Safety Evaluation in Robotic Manipulation
Chengyue Huang, Khang Vo Huynh, Sebastian Elbaum, Zsolt Kira, Lu Feng · · 2026
Summary
Robotic manipulation is typically evaluated by task success, but successful completion does not guarantee safe execution. Many rl failures are temporal: a robot may touch a clean surface after contamination or release an object before it is fully inside an enclosure.
Abstract Summary
Key Points
- Introduces SafeManip, advancing simulation research.
- Detailed simulation demonstrates effectiveness on benchmarks and real-world scenarios.
- Provides insights into methodology and architectural choices.
Temporal Safety Properties for Robotic Manipulation
A central contribution of SafeManip is to make manipulation safety formally specified, reusable across tasks, and temporally checkable. Many safety failures are not defined by a single unsafe state, but by how events unfold over time: whether an object remains stable after grasping, or whether contamination from handling raw food is followed by sanitization before clean contact.
This section formalizes such failures as temporal safety properties.
Benchmark Implementation and Evaluation Protocol
We release a repository 1 1 1 Code URL: https://github.com/chengyuehuang511/SafeManip containing the SafeManip property templates, task bindings, temporal monitors, rollout processing code, and evaluation scripts.
We instantiate SafeManip in RoboCasa [ 13 ] , a kitchen manipulation simulator that provides articulated fixtures, appliances, diverse objects, and multi-step household tasks while exposing simulator state for grounding symbolic safety predicates.
Experimental Results
We report experimental results from evaluating the six VLA policies and training variants described in Section 4.4 on the 50 RoboCasa tasks introduced in Section 4.1 . Each evaluated policy is run for 50 rollouts per task, and every rollout is monitored using the instantiated temporal safety properties defined in Section 3 and the monitoring pipeline described in Section 4.2 .
This protocol produces the task completion, temporal safety violation, rollout outcome, and unsafe-state exposure metrics described in Section 4.3 .
Sources and Demonstrations
Figures are reproduced from Huang et al. See the full paper for experimental details and the project page, when the paper links one, for demonstrations and videos.
Related Papers
EgoForce: Forearm-Guided Camera-Space 3D Hand Pose from a Monocular Egocentric Camera
Christen Millerdurai, Shaoxiang Wang, Yaxu Xie et al.
Reconstructing the absolute 3D pose and shape of the hands from the user's viewpoint using a single head-mounted camera is crucial for practical egocentric interaction in AR/VR, telepresence, and hand-centric manipulation tasks, where sensing must remain compact and unobtrusive.
CoRe: Combined Rewards with Vision-Language Model Feedback for Preference-Aligned Reinforcement Learning
Hexian Ni, Tao Lu, Yinghao Cai
CoRe combines code-based rewards with learned visual preferences so that neither reward source has to capture the entire task alone.
Aligning Flow Map Policies with Optimal Q-Guidance
Christos Ziakas, Alessandra Russo, Avishek Joey Bose
Generative policies based on expressive model classes, such as diffusion-models and flow matching, are well-suited to complex control problems with highly multimodal action distributions.
Human-Centric Transferable Tactile Pre-Training for Dexterous Robotic Manipulation
Chi Zhang, Penglin Cai, Ziheng Xi et al.
TTP transfers tactile experience from human demonstrations into robot policies for contact-sensitive manipulation.