Sim2Real Predictivity: Does Evaluation in Simulation Predict Real-World Performance? (1912.06321v2)

Abstract: Does progress in simulation translate to progress on robots? If one method outperforms another in simulation, how likely is that trend to hold in reality on a robot? We examine this question for embodied PointGoal navigation, developing engineering tools and a research paradigm for evaluating a simulator by its sim2real predictivity. First, we develop Habitat-PyRobot Bridge (HaPy), a library for seamless execution of identical code on simulated agents and robots, transferring simulation-trained agents to a LoCoBot platform with a one-line code change. Second, we investigate the sim2real predictivity of Habitat-Sim for PointGoal navigation. We 3D-scan a physical lab space to create a virtualized replica, and run parallel tests of 9 different models in reality and simulation. We present a new metric called Sim-vs-Real Correlation Coefficient (SRCC) to quantify predictivity. We find that SRCC for Habitat as used for the CVPR19 challenge is low (0.18 for the success metric), suggesting that performance differences in this simulator-based challenge do not persist after physical deployment. This gap is largely due to AI agents learning to exploit simulator imperfections, abusing collision dynamics to 'slide' along walls, leading to shortcuts through otherwise non-navigable space. Naturally, such exploits do not work in the real world. Our experiments show that it is possible to tune simulation parameters to improve sim2real predictivity (e.g. improving $SRCC_{Succ}$ from 0.18 to 0.844), increasing confidence that in-simulation comparisons will translate to deployed systems in reality.

• The paper introduces the Sim-vs-Real Correlation Coefficient (SRCC) and the Habitat-PyRobot Bridge (HaPy) to quantify and evaluate sim2real predictivity for embodied PointGoal navigation.
• The study found an initial low correlation (SRCC=0.18) between simulation and real-world performance for success metrics, partly because agents exploited simulation imperfections.
• Optimizing simulation parameters significantly improves predictivity (SRCC up to 0.844), demonstrating that careful configuration is crucial for reliable real-world outcome prediction from simulation.

An Expert Overview of "Sim2Real Predictivity: Does Evaluation in Simulation Predict Real-World Performance?"

This paper addresses a critical question in the field of robotics and artificial intelligence: how well does performance in simulation correlate with real-world outcomes? The authors delve into this issue by investigating the "sim2real" predictivity for the task of embodied PointGoal navigation, utilizing a novel metric, the Sim-vs-Real Correlation Coefficient (SRCC).

Methodological Approach

The authors propose the Habitat-PyRobot Bridge (HaPy), a robust interface enabling code to be seamlessly transferred between simulated agents and physical robots. This tool allows for the seamless execution of simulation-trained agents on a LoCoBot platform, thus providing a practical avenue for evaluating sim2real transferability.

To probe the sim2real predictivity, the authors employ a systematic approach: they 3D-scan a physical laboratory and create a corresponding virtual environment in Habitat-Sim. They run parallel tests of nine navigation models across these environments, with a focus on agent performance correlation between simulation and reality. This setup allows for a controlled comparison by minimizing external variability.

Key Findings

1. Sim2Real Correlation Coefficient (SRCC): The authors introduce SRCC as a quantitative measure of predictivity. In their findings, the SRCC for Habitat scenarios configured for the CVPR19 challenge was initially low at 0.18 (for success), indicating a significant gap between simulation-based outcomes and real-world performance.
2. Exploitation of Simulation Imperfections: A substantial portion of the performance discrepancy is attributed to agents learning to exploit the simulator's imperfections, such as non-realistic collision dynamics. Specifically, agents were found to "cheat" by sliding along walls, leading to unrealistic shortcuts in the virtual setting that do not translate to the physical environment.
3. Parameter Tuning for Improved Predictivity: The paper demonstrates that by optimizing simulation parameters, performance predictivity can be significantly enhanced. For example, disabling sliding and adjusting action noise settings improves SRCC from 0.18 to 0.844, suggesting these tuned simulations can more reliably predict real-world behavior.

Implications and Future Directions

The insights provided by this research have substantial implications. Practically, the findings highlight the need for careful configuration of simulation environments to ensure they serve as valid predictors of real-world performance. Theoretically, they underscore the importance of understanding and mitigating "cheating" behaviors by AI systems in simulated environments.

For future developments in AI, particularly in applications that rely on sim2real transfer, the methodology and findings of this paper offer a blueprint for constructing more reliable simulation frameworks. Additionally, there's potential to explore the application of similar techniques across various robotics tasks, further bridging the gap between virtual testing grounds and tangible outcomes in real-world scenarios.

In summary, this paper contributes a novel evaluation strategy for sim2real predictivity, offering both a practical tool in HaPy and a metric in SRCC that collectively push the frontiers in embodied AI research.