Full Papers
Up to 4 pages using the CoRL 2026 anonymized template; non-archival; accepted papers presented as posters, with top submissions invited to give lightning talks.
Machine Learning/CORE Unranked
SPAIS CoRL 2026
The Science of Physical AI Safety workshop @ CoRL 2026
The SPAIS 2026 workshop, held in conjunction with CoRL 2026 in Austin, Texas, brings together AI safety and robot learning communities to establish a shared scientific foundation for the interpretability, alignment, and control of robot foundation models (RFMs). It focuses on three open questions: whether AI-safety techniques from LLMs transfer to RFMs, whether classical robotics safety methods apply, and whether new evaluation methods are needed for embodied agents.
Paper fit
Contribution paths
A strong submission should clearly identify its contribution and evaluate it appropriately.
Demo Submissions
Visual media plus optional short analysis of an RFM safety failure beyond collision; selected demos are spotlighted during the workshop.
Research areas in scope
01
Interpretability
Probing, sparse autoencoders, activation patching, and steering applied to RFMsLocalizing safety-relevant concepts in model internals (e.g. human proximity, fragile objects, force limits, task boundaries)What changes when interpretability moves from text to continuous, closed-loop action, and how to ground internal findings in physical outcomes
02
Alignment & Robustness
Specifying and instilling safe behavior: robot constitutions, RL and supervised fine-tuning, guardrailsWhether LLM safety training survives the transfer to embodimentJailbreaks, prompt injection, and adversarial attacks across the embodied attack surface (observations, instructions, and the physical environment)Defenses against embodied adversarial attacks
03
Control
Runtime monitoring, anomaly detection, and competence estimation for deployed policiesExtending classical guarantees (reachability analysis, safety filters, constrained control) to learned policies and latent spacesOversight of embodied agents: monitoring, sandboxing, and shutdown in physical environmentsUnderstanding and improving sim-to-real transfer of safety propertiesSecurity of deployed robot fleets
04
Evaluation
Benchmarks, datasets, red-teaming, and hardware-in-the-loop protocols for RFM safety beyond collision avoidanceEvaluation methods for open-ended, general-purpose physical systems
Policies worth checking twice
- Submissions are double-blind reviewed.
- Papers must be up to 4 pages using the CoRL 2026 template.
- Submissions are non-archival.
- Demonstrations are accepted separately from papers and have their own submission process.
- Travel grant applications are reviewed independently of paper submissions.
Official sources
Compiled from the official call for papers. The organizers’ pages remain authoritative.
Official workshop siteCall for Papers CFP Papers (up to 4 pages) on interpretability, alignment, contr
Last verified September 9, 2026