Explore an episode
Switch cameras, move through decisions, and inspect generated annotations beside the footage.
Open the studio ↗Datumwell / Embodied-agent data / v1.0
EmbodiedTrace
See the task.
Explore the subgoal.
Read the action.
Visual observations, task-oriented annotations, and structured robot actions — together in an inspectable 100-episode sample.
For teams exploring data representations for embodied agents.
100episodes to explore
4,081structured action decisions
1,288generated subgoal segments
3synchronized camera views
A concrete starting point
Evaluate the representation.
Then discuss the fit.
Start with the example. Inspect the observations, read the generated subgoal, and compare the structured action with your own interface.
Task
Put the marker in the pot
Subgoal · generated
Approach / align with marker
Recorded-command action label · abridged
move_eef(position_m=[0.432, 0.109, 0.481],
quaternion_xyzw=[−0.988, −0.034, −0.117, 0.091],
gripper_closure=0.0)An end-effector (EEF) pose action describes a target position and orientation for the robot's working end, alongside a gripper command. This example is rounded for display; the dataset retains full precision.
Three ways to inspect
From first look
to integration.
Inspect the format
Load one decision before downloading media. The dataset card includes a short Python example.
Read the dataset card ↗Bring your requirements
Discuss your task coverage, observation inputs, action representation, and annotation needs.
Discuss a pilot ↗Your tasks. Your action format.
What would a useful pilot
look like for your team?
Share the problem you're working on. We can discuss a scope, sample format, and evaluation criteria before any engagement.
Discuss your data requirements ↗Prepare a pilot brief → A checklist to share with your team.
Useful starting points for the conversation
- Tasks and environments you care about
- Available observations and action interface
- Required annotations and review criteria
- Sample deliverables and intended evaluation
Before you evaluate
A few practical
questions.
What is generated in this sample?
Subgoals, concise rationales, and action descriptions are generated labels for inspection. They are not recorded human thoughts or verified task completion. Original observations and recorded-command action labels remain distinguishable in the package.
Can I use these actions directly on my robot?
The sample uses robot-base end-effector pose targets. Your team should assess embodiment compatibility and an appropriate action adapter. The derived sparse targets have not been separately executed; no robot-performance improvement is claimed.
What do the downloads contain?
The dataset package contains 4,081 decision rows across episode-level splits, a schema, and optional original JPEG/video archives. The studio's per-episode JSON download is an inspection export, rather than the training package.
Does this demo cover our production task distribution?
The 100 episodes provide a representation sample, not a claim of production coverage. Read the coverage summary and bring your required tasks, inputs, and annotation criteria to a scoping discussion.