HomeIndustriesNvidia's DrEureka outperforms humans in training robotic systems

Nvidia's DrEureka outperforms humans in training robotic systems

According to a, large language models (LLMs) can speed up the training of robotic systems in superhuman ways recent study by scientists NvidiaThe University of Pennsylvania and that University of Texas, Austin.

The study presents DrEureka, a way that may routinely create reward functions and randomization distributions for robotic systems. DrEureka stands for Domain Randomization Eureka. Requiring only a high-level description of the goal task, DrEureka is quicker and more efficient at transferring learned policies from simulated environments to the true world than human-developed rewards.

The implications will be enormous for the fast-moving world of robotics, which has recently received a lift from advances in speech and vision models.

Sim-to-Real Transfer

When designing robotics models for brand spanking new tasks, a policy is usually trained in a simulated environment and deployed in the true world. The difference between simulation and real environments, called the “sim-to-real” gap, is one in every of the most important challenges of any robotics system. Configuring and fine-tuning the policy for optimal performance typically requires some forwards and backwards between simulation and real-world environments.

Recent work has shown that LLMs can mix their extensive world knowledge and reasoning skills with the physics engines of virtual simulators to learn complex, low-level skills. For example, LLMs will be used to design reward functions, the components that control the robotic reinforcement learning (RL) system, to seek out the right motion sequences for the specified task.

However, once a policy is learned in simulation, transferring it to the true world requires many manual adjustments to the reward functions and simulation parameters.


The goal of DrEureka is to make use of LLMs to automate the intensive human efforts required for the transfer process from simulation to reality.

DrEureka is built on Eureka, a technology launched in October 2023. Eureka takes a robot task description and uses an LLM to generate software implementations for a reward function that measures success on that task. These reward functions are then executed within the simulation and the outcomes are returned to the LLM, which reflects the result and converts it into the reward function. The advantage of this method is that it will possibly be run in parallel with a whole bunch of reward functions, all generated by the LLM. It can then select the perfect features and improve them further.

While Eureka's reward functions are great for training RL policies in simulation, they don’t consider the clutter of the true world and due to this fact require manual translation from simulation to reality. DrEureka addresses this shortcoming by routinely configuring domain randomization (DR) parameters.

DR techniques randomize the physical parameters of the simulation environment in order that the RL policy will be generalized to the unpredictable disturbances encountered in the true world. One of the important thing challenges of DR is choosing the correct parameters and the correct perturbation range. Adjusting parameters requires sound physical pondering and knowledge of the goal robot.

“These features of DR parameter design make it a really perfect problem for LLMs, as they’ve strong physics knowledge and effective hypothesis generation, and supply good initializations for complex zero-shot search and black-box optimization problems .” the researchers wrote.

DrEureka uses a multi-step process to interrupt down the complexity of concurrently optimizing reward functions and domain randomization parameters. First, an LLM generates reward functions based on a task description and safety instructions in regards to the robot and the environment. DrEureka uses these instructions to create an initial reward function and learn a policy like in the unique Eureka. The model then runs tests on the policy and reward function to find out the suitable range of physical parameters similar to friction and gravity.

The LLM then uses this information to pick the optimal domain randomization configurations. Finally, the policy is retrained with the DR configurations to grow to be more robust to real-world noise.

The researchers described DrEureka as “a language model-driven pipeline for simulation-to-reality transfer with minimal human intervention.”

DrEureka in motion

The researchers evaluated DrEureka on four-legged and dexterous manipulator platforms, although the strategy is general and applicable to different robots and tasks. Their results show that DrEureka-trained four-legged locomotion guidelines outperform classical human-developed systems by 34% in forward speed and 20% in distance traveled across various real-world evaluation terrains. They also tested DrEureka for skillful manipulation with robotic hands. For a set period of time, the perfect policy trained by DrEureka performed 300% more cube rotations than human-developed policies.

The most interesting result, nonetheless, was the appliance of DrEureka to the novel task of balancing and walking a robo-dog on a yoga ball. The LLM was capable of design a reward function and DR configurations that allowed the trained policy to be transferred to the true world without additional configurations and to perform adequately on various indoor and outdoor terrains with minimal security support.

Interestingly, the study found that the security instructions contained within the task description play a vital role in ensuring that the LLM generates logical instructions which might be transferred to the true world.

“We imagine that DrEureka demonstrates the potential to speed up robot learning research by utilizing foundational models to automate the difficult design features of low-level skill learning,” the researchers write.


Please enter your comment!
Please enter your name here

Must Read