Eric Jang: The Imperative of Ephemeralizing Robots in the Modern Age.

**Two Approaches to Sim2Real Robotics Research**
There are two main approaches to sim-to-real robotics research. One approach involves creating an “adapter” that transforms simulated sensor readings to resemble real data, allowing a robot trained in simulation to perform the same tasks in the real world. Another approach involves training simulated robots under various randomized conditions, making them robust enough to handle different scenarios in the real world.

**The Case for Direct Iteration in the Real World**
In the past, I was skeptical about sim2real robotics research and believed that real-world data was essential for studying general-purpose robots. I argued that the complexity and messiness of the real world provided invaluable learning experiences that simulations couldn’t replicate. Additionally, bringing real-world objects and interactions into a simulation was challenging and often unrealistic. I also questioned the effectiveness of bridging the “reality gap” with high-dimensional generative models.

**Lessons Learned and Changing Perspectives**
However, my perspective shifted after working on a successful three-year robotics project that focused on direct iteration in the real world. Through this experience, I realized the challenges in evaluating general-purpose robots and the need for offline evaluation techniques like simulation. I now believe that offline evaluation is essential for studying general-purpose robots and have adjusted my research workflows accordingly.

**The Problem of Success: Evaluating General-Purpose Robots**
The emergence of general-purpose robots that can perform a wide range of tasks poses new challenges in robotics research. Evaluating such robots becomes increasingly difficult when their success rate is moderate, and they need to generalize to thousands or even millions of operating conditions. The initial excitement of a robot learning numerous tasks quickly turns into uncertainty about how to improve the system further.

**Moving Towards Sim2Real Evaluation**
To address the challenges of evaluating general-purpose robots, sim2real evaluation techniques become crucial. By iteratively training and evaluating robots in simulation, researchers can fine-tune their models before deploying them in the real world. This approach allows for faster iterations, reduces operational complexity, and enables the evaluation of policies without the need for expensive and time-consuming real-world experiments.

**Benefits and Limitations of Sim2Real Evaluation**
Sim2Real evaluation techniques offer several advantages for robotics research. Firstly, simulation provides statistical reproducibility and facilitates learning without safety concerns. Additionally, simulations allow for the generation of new datasets with every code change, making it easier to adapt and iterate. Moreover, sim2real evaluation can be applied to hardware platforms that have limited trials, preventing excessive wear and tear on the robot.

However, sim2real evaluation also has its limitations. Simulations may not perfectly represent the real world, and the process of domain adaptation or domain randomization is still an ongoing research problem. It is crucial to ensure that the policies learned in simulation can be effectively transferred to the real world.

**Conclusion**
In conclusion, while I was initially skeptical of sim2real robotics research, my perspective has changed based on the challenges of evaluating general-purpose robots. Sim2Real evaluation techniques, such as training and testing robots in simulation, offer significant benefits in terms of faster iterations, reproducibility, and reduced complexity. While sim2real evaluation is not without its limitations, it is becoming increasingly essential in the field of robotics research. By embracing sim2real techniques, researchers can push the boundaries of what robots can achieve in the real world.