Overview
Workshop Description
A community-wide effort to unify simulation benchmarking and real-world robot deployment.
Simulation has become the engine of modern embodied AI. Physics-grounded environments
like BEHAVIOR provide the scale, safety, and reproducibility that physical data collection
cannot, enabling agents to develop high-level reasoning, long-horizon planning, and
dexterous bimanual manipulation across thousands of everyday scenarios. Rigorous diagnostic
frameworks like RoboEval further deepen what we can learn from simulation by instrumenting
every task with stage-level metrics that reveal not just whether a policy succeeds, but
precisely how and why it fails. Yet evaluation alone is not enough: RoboChess puts
simulation-trained policies to the ultimate test by requiring them to transfer directly
to physical robots, perceiving a real chessboard, planning long-horizon move sequences,
and executing dexterous manipulation across multiple hardware embodiments. The frontier
question is therefore not whether to train in simulation, but how to ensure that the
capabilities learned at scale in simulation carry over robustly to the physical world.
This workshop unites benchmark designers, robot-learning researchers, foundation-model
developers, and real-robot practitioners to address that challenge together.
Large-Scale
Simulation TrainingGPU-Accelerated Policy Learning
→
Rigorous Benchmark
EvaluationStandardized Metrics & Reproducibility
→
Real-World
TransferRobust Deployment & Generalization
Research Questions
Core Challenges
🏠 Simulation as the Best Training Ground
Rich, physics-grounded environments enable agents to practice thousands of household tasks
at scale — from high-level reasoning and long-range navigation to dexterous bimanual
manipulation — far beyond what physical data collection alone can provide.
🔬 Beyond Binary Success: Diagnostic Evaluation
Task completion alone is not enough. Stage-level metrics that measure outcome, efficiency,
bimanual coordination, and safety reveal precisely where and why policies fail, enabling
targeted improvement rather than opaque leaderboard scores.
🎯 From Demonstration to True Task Solving
Many policies succeed on curated demonstrations but fail to generalize to diverse initial
states and long-horizon settings. Closing this gap requires evaluation protocols that
stress-test policies across a wide range of conditions in simulation before real-world deployment.
🚀 Sim-to-Real Deployment
Translating simulation-trained policies to physical robots remains an open challenge.
Robustness to lighting, object position, sensor noise, and embodiment differences must
be systematically evaluated so that strong simulation performance predicts real-world capability.
Embodied AI Challenge Suite
The workshop is organized around the CoRL 2026 Embodied AI Benchmark and Sim-to-Real Challenge Suite,
consisting of three complementary tracks. Each track retains its canonical technical infrastructure,
evaluation protocol, leaderboard, and prizes. Across tracks, participants will report a shared set of
outcome, efficiency, robustness, safety, computational-cost, and reproducibility measures, synthesized
into a community white paper.
