DROID
Large-scale, in-the-wild robot manipulation demonstrations collected across many environments and tasks.
About
DROID brings together diverse robot demonstrations recorded across different scenes, operators, and manipulation tasks. That variety makes it useful for studying policies that must generalize beyond one carefully controlled lab setup.
The most productive way to approach a corpus at this scale is to begin with a small, inspectable cohort. Use the live Rerun view to understand how observations, actions, and timing line up before building a broader preprocessing or training job.
Suggested agent workflows
Copy a prompt into your agent of choice. Each workflow starts small, respects the source terms, and makes the result inspectable in Rerun.
Audit a training slice
Fetch DROID from its official source at https://droid-dataset.github.io/. Read and preserve its source terms and citation requirements. Build a reproducible manifest for a representative slice, then load it into Rerun. Inspect the RGB video, Robot state, Actions streams for timestamp gaps, missing samples, calibration problems, and annotation inconsistencies. Summarize the findings and save both the validation script and a Rerun recording that makes every issue easy to review.
Build a baseline pipeline
Use DROID to scaffold a baseline pipeline for Imitation learning, Manipulation. Start with a small deterministic cohort and document the selection criteria. Create the dataset adapter, preprocessing steps, train and validation splits, and one lightweight baseline. Log inputs, labels, predictions, losses, and representative failures to Rerun so I can inspect the full experiment timeline. Include exact setup and reproduction commands.
Find useful failure cohorts
Analyze DROID and propose a practical cohort taxonomy for Imitation learning, Manipulation. Use available metadata and the RGB video, Robot state, Actions signals to identify hard cases, rare conditions, coverage gaps, and possible leakage between splits. Materialize the cohorts as queryable metadata, log examples from each cohort to Rerun, and produce a short report explaining which cohorts should become regression tests. Do not infer labels that are not present in the source data.