Fast experiment loops for robotics: simulation, evaluation, and dataset generation

Jobs suit workloads with a defined end, such as dataset conversion, policy fine-tuning and batch evaluation. Endpoints keep a service available for requests, such as a model that returns predictions or robot actions through an API.

In the Positronic workflow, data conversion and training run as jobs, while policy inference runs through an endpoint. This separates the resources needed to produce a checkpoint from those needed to test it on a robot.

Package the training environment in a Docker image, configure the dataset and policy, and submit the training command as a serverless GPU job. The session walks through examples for ACT and diffusion policies using LeRobot.

The examples save checkpoints to S3 storage and support experiment tracking through Weights & Biases. The presenters explain the configuration and commands using prepared examples; a complete training run is not performed during the webinar.

Robotics models can expect different dataset formats and expose different inference interfaces. Comparing them requires integration work even when they are trained on demonstrations of the same task.

Positronic Robotics describes separate data conversion steps for its model implementations and a shared inference API around them. This allows its evaluation setup to switch models without changing the robot-side interface. The team also keeps the operator unaware of which model is running during a benchmark episode.

Yes. In the live demonstration, a robot in Positronic’s lab connects to a policy served on Nebius. The remote model returns chunks of actions for the robot to execute locally.

The speaker reports roughly 300 milliseconds of round-trip latency, including inference, and action chunks covering around 1.5 to 2 seconds. These figures describe this setup, not a general latency guarantee. The implementation still pauses between chunks, so teams must assess whether remote inference fits their task’s timing requirements.

Inspect the model output, runtime logs and connection settings to distinguish a policy failure from an infrastructure problem. Keeping the configuration and checkpoint associated with each run helps teams investigate what happened.

The webinar shows both kinds of issue. A simplified vision-language model demo violates an obstacle constraint, prompting discussion of instructions and model choice. The real-robot demo initially times out because it uses an old endpoint IP address; correcting the address allows the robot to connect.

Simulation rollouts, synthetic dataset generation and policy sweeps are potential workloads for independent batch jobs. Teams can vary conditions or configurations, collect outputs and use the results to guide further training and evaluation.

The webinar introduces these as use cases for serverless infrastructure. Its detailed examples focus on policy fine-tuning, dataset conversion and inference. Simulation results still need validation against the physical task before teams can draw conclusions about real-robot performance.