← Back to Praxis
Praxis — live manipulation sample
A real episode, annotated phase by phase. Watch action, grasp, contact, and failure update live as the clip plays.
t = 0.00 s✓ stable
Schema
v0.1 · ~20 fields / segment · 8+ / hand
Export
LeRobot parquet / JSON
Typical label
Captures actions and outcomes.
task: "pick up the towel and place it in the target bin"
segments: ["reach", "grasp", "lift", "grasp", "place"]
success: 1
Phase names only. No indication of which grasp was the recovery, no slip signal, no per-phase outcome.
Praxis label · start
segment 0.00–1.00s
Captures transitions, failure, and recovery.
action:"reach"
object:"towel"
language:"reach for the towel"
object_state_start:"in_source_bin"
object_state_end:"in_source_bin"
outcome:"success"
right:
role:"primary"
grasp_type:null
grasp_quality:null
grasp_location:null
contact_state:"no_contact"
Praxis captures structured action sequences from real manipulation episodes across tasks, objects, and environments.
Derived from motion and interaction signals, not language descriptions.
Real-world episodes
First-person recordings across diverse tasks and environments.
Failure is included
Captures slips, retries, and adjustments, not just successful outcomes.
Consistent across tasks
The same representation applies across different objects and manipulation scenarios.
Training-ready output
Structured, time-aligned data designed for learning policies, not just perception models.
Built for systems operating outside controlled environments.