tasks int64 | native_outcomes_by_family dict | usage_complete_records int64 | verified_videos int64 | native_hd_videos int64 | original_observation_clips int64 | native_instructions int64 | modified_instructions int64 | recording_incomplete_count int64 | instruction_revision_complete bool | completed_instruction_revisions int64 | pending_instruction_revisions int64 |
|---|---|---|---|---|---|---|---|---|---|---|---|
31 | {
"task01": {
"tasks": 8,
"success": 1
},
"task02": {
"tasks": 2,
"success": 2
},
"task03": {
"tasks": 5,
"success": 4
},
"task04": {
"tasks": 8,
"success": 4
},
"task05": {
"tasks": 8,
"success": 1
}
} | 31 | 31 | 31 | 0 | 13 | 18 | 0 | true | 15 | 0 |
RLE-Bench · Kinex
All eight authorized completion replacements are published with complete native evidence, provider usage and verified videos. All fifteen instruction revisions now have a normally finished current result. The task04/08 recording-fix replacement uses Kinex 0.9.0, Astra medium and seed 0; it is valid and unsuccessful, stopped by the agent after 1,675 native action frames. Its 60 requests have complete terminal usage. No new model attempt was launched for this publication.
The current selection contains 31 tasks across five environments. Native successes remain RoboCasa 1/8, LIBERO 2/2, RoboTwin 4/5, RoboDojo 4/8 and BEHAVIOR 1/8. These are separate single-seed family observations, not a pooled benchmark estimate. All 31 current results are normally finished and usage-complete. Both LIBERO results retain their measured Kinex 0.8.0 identity.
The new pour video restores recorded scene and liquid-surface states, with verified intermediate states and no command resimulation. It uses two cameras, 4x playback and 20 fps. Other episodes retain their original replay validation scope. Accepted recorded-state views confirm the tipped violet bottle and displaced white bowl; they do not by themselves establish a controller-level failure cause.
Prior unsuccessful, interrupted, cancelled and superseded attempts remain indexed and preserved. Historical request-bearing attempt counts remain lower bounds where cancelled logs are incomplete; current evaluation completion does not claim complete accounting for every historical interruption. Existing egg-seating calibration questions remain unresolved. Native criteria, task selection and budgets are unchanged.
Instruction audit · Attempt history · Execution audit · Video inventory
This edition uses an archived offline-only USD pose synchronization patch for task04/08. All 1,676 physics and fluid-surface states are checked; rigid USD poses are checked against each restored state, with zero physics steps advanced. Initial, middle and terminal camera images are visually reviewed. The original run, native verdict and recorded state files are unchanged.
- Downloads last month
- 3,872