220 episodes · 20 fps · 2 cameras · 224×224 h264
MemCert · batteries_sort · L2 · 224 px
Part of MemCert, a certified robot-memory benchmark: simulated Panda manipulation episodes where the correct action depends on something the robot saw earlier and can no longer see. Each level is certified so that no observation channel at decision time leaks the hidden answer, and (for memory sets) so that the answer is recoverable from history.
Contents
| Format | LeRobot v3.0, robot panda, 20 Hz |
| Episodes | 220 (train 200, val 20) |
| Frames | 431,505 (6.0 h) |
| Cameras | observation.images.top 224×224, observation.images.wrist 224×224 |
| Episode length | mean 1961, max 3268 steps |
| Expert success (raw; readings are noisy by design) | 182/220 |
| Level dials | count=4, occlusion=0.15, flip=0.05, spill=0.2, interference=1 |
Per-episode metadata (seed, split, dials, success) is in meta/memcert_episodes.jsonl. Frame-level labels
(annotation.mode, annotation.lamp_on, annotation.lamp_shown, annotation.lamp_occluded, annotation.reading_flip, annotation.spill_event, …) are for analysis only and must not be given to a policy as input.
Certificate
- Verdict: pass (protocol
cv5, chance 0.5) - Leak probes, two-sided AUROC (gate ≤ 0.60): top 0.502, wrist 0.523, proprio 0.512, step 0.513, language 0.500
- Memory is needed: history AUROC 0.899 (gate ≥ 0.85)
- Expert motion leak, two-sided AUROC 0.512 (gate ≤ 0.60)
- Decisions probed: 880
Full numbers: meta/certificate.json.
Load
from lerobot.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("hiwe0305/MemCert-BatteriesSort-L2")
License: CC BY 4.0.
- Downloads last month
- 106