CrossSafe: Towards Cross-Embodiment Latent Safety Filters
Ihab Tabbara, Yuxuan Yang, Hussein Sibai
cs.RO · cs.AI · cs.LG · eess.SY
Quick verdict
Proposes a shared HJ safety filter generalizing zero-shot across 5 robot embodiments; promising sim results but high residual collisions and muddled ablations.
Summary
The paper proposes CrossSafe, an embodiment-conditioned latent safety filter in which a single Hamilton-Jacobi reachability critic and safety-maximizing policy are jointly trained across heterogeneous robot embodiments. Building on a frozen HoloBrain-0 encoder, the authors augment per-link robot tokens with safety features (link velocities, orientations, kinematic-tree depth), add geometry-aware manipulator-scene cross-attention biased by 3D distance and direction, and let the actor emit per-joint displacements scaled by URDF velocity limits, supporting arms with different degrees of freedom. Training uses SAC adapted to HJ reachability in RoboTwin 2.0 with five bimanual embodiments and five manipulation tasks under a static obstacle-avoidance constraint, evaluated via five leave-one-embodiment-out folds (5,000 in-distribution and 1,250 out-of-distribution episodes per method). CrossSafe lowers the nominal planner's collision rate from 64.1% to 39.8% (in-distribution) and 49.8% (OOD) and contact force from 179.6 N to 102.9 N (OOD), generalizing zero-shot to held-out embodiments, including transfer from 6-DoF training arms to a 7-DoF Franka-Panda. Main limitations: simulation-only with one static obstacle and a nominal planner deliberately blind to it; residual collision rates remain high (40-50%); in-distribution task success drops (32.2% to 23.7%); ablation evidence for the feature augmentation is mixed (w/o Aug slightly beats the full model on OOD collision rate, 49.7% vs 49.8%); no significance testing; and 73 ms inference exceeds the 40 ms control period, leaving real-time deployment unaddressed.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 4
● low 1
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A single HJ reachability critic and safety-maximizing policy can be jointly trained across four embodiments and generalize zero-shot to a fifth unseen embodiment, including transfer from 6-DoF training arms to a 7-DoF Franka-Panda (trained components never see 7-DoF data).
- CrossSafe reduces the nominal collision rate from 64.1% to 39.8% InD and 49.8% OOD and contact force from 179.6 N to 102.9 N OOD, while intervening on only 9.2-13.2% of control steps.
- Training the safety filter across more embodiments improves OOD outcomes relative to single-embodiment specialists (CR 49.8% vs 54.4%; Force 102.9 N vs 144.8 N).
- Geometry-aware attention and safety-feature augmentation are the components that enable OOD generalization, evidenced by all three ablations exceeding the nominal CR in the held-out 7-DoF Franka fold.
- CrossSafe dominates the X-VLA-Safe variant on all four metrics in both InD and OOD evaluations (e.g., InD CR 39.8% vs 50.4%; SR 23.7% vs 17.0%).
Strengths
Novel problem formulation with variable-DoF architecture
First jointly trained cross-embodiment HJ safety critic and safe policy (Sec. I contributions); the shared per-link linear Gaussian actor head with tanh-squashing and per-joint URDF velocity-limit scaling (Sec. IV-D, App. B-2) is what enables the demonstrated zero-shot transfer from four 6-DoF embodiments to a 7-DoF Franka-Panda (Sec. VI-b: CR 66.0% to 60.4%, SR 40.4% to 45.2%).
Unusually rigorous evaluation protocol
Five leave-one-embodiment-out folds, deterministic shared scenes with re-run reproducibility (App. C-1), 50 episodes per task-embodiment cell, 5,000 InD plus 1,250 OOD episodes per method, full per-pair breakdowns (Tables V, VI), plus specialists-vs-generalists (Table II) and a second-encoder instantiation (X-VLA-Safe, Sec. V-D).
Honest negative reporting
Well-motivated geometry-aware design
Deployment-relevant engineering data
Weaknesses
Manufactured baseline; most-informative comparison absent
Residual safety and success levels are weak
Ablation evidence contradicts the design story on the headline metric
No inferential statistics
Real-time feasibility gap and simulation-only scope
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
CrossSafe: Towards Cross-Embodiment Latent Safety Filters — Preliminary Review
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is CrossSafe and what problem does it solve?
CrossSafe is a learned safety filter for robot manipulators: a single Hamilton-Jacobi reachability value function plus a safety-maximizing policy, trained jointly across several different robot arms so that it can also protect robot embodiments it was never trained on. A nominal task policy proposes actions; the filter evaluates each action's safety value and swaps in a safe joint-space action when the value is negative.
How much does CrossSafe reduce collisions?
In RoboTwin 2.0 simulation with a static box obstacle placed on the robot's pick-and-place path, the unfiltered planner collided in 64.1% of episodes. CrossSafe cut this to 39.8% on training embodiments and 49.8% on a held-out unseen embodiment, and lowered measured contact force from 179.6 N to about 103-114 N. Collisions remain frequent, so the result is a proof of concept rather than a deployable safety guarantee.
Does CrossSafe work on robots with different numbers of joints?
Yes — that is the core design goal. The actor emits one displacement per joint from shared per-link modules, scaled by each joint's velocity limit from the robot's URDF, so the architecture is agnostic to degrees of freedom. In the strongest result, a filter trained only on 6-DoF arms transferred zero-shot to a 7-DoF Franka-Panda, reducing its collision rate from 66.0% to 60.4% and raising task success from 40.4% to 45.2%.
Is code and data available for CrossSafe, and can it run on real robots?
The paper states that videos and code are available at the project page (trustworthyautonomy.github.io/CrossSafe) and gives complete hyperparameters on the public RoboTwin 2.0 simulator, with deterministic evaluation scenes. However, all experiments are simulation-only, use a single static obstacle, and inference (73 ms) currently exceeds the 40 ms control period, so hardware validation remains future work.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-2 are architecture/task overview diagrams recovered only as fragmentary text; visual layout and attention-map renderings could not be verified from the extraction.
- Equations 2 and 4-6 are partially garbled by text extraction; interpretations rely on the surrounding prose, which is consistent.
- The claimed code availability at the project page could not be verified (no network access in this review).
- Appendix Table III/IV contents are present, but some appendix subsection headers appear as table-of-contents artifacts; minor implementation details may be missing.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.