YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
FlyJEPA embodied Drosophila corpus (v2)
A benchmark of whole-brain connectome-driven embodied behaviour. See the task tiering before reading any aggregate number.
What this is
Rollouts of a whole-brain Drosophila connectome leaky-integrate-and-fire (LIF) model embodied in NeuroMechFly. 945 episodes / 142689 decision windows. The brain is driven by afferent Poisson input from the body's senses and its descending readout is decoded back into a body command each 50 ms window.
Task tiering (8 tasks + 1 negative-data control, NOT 9 flat tasks)
- Taxis & discrimination (4): sugar_approach, good_vs_bad, strong_vs_weak, scent_trail.
- Navigation (3): gauntlet, obstacle_field, meadow.
- Controls (2): aversive_avoid (a reflex probe) and explore (stimulus-only). These are controls / negative data, not solvable tasks - do not count them as solved competence.
The explore control is a per-episode coin flip
55/105 explore episodes carry a zero-intensity placeholder (no cue at all); 50/105
have a real weak appetitive source. The difficulty knob only applies to the sourced
half, and the sourced/unsourced split is RANDOM, not balanced per cell. Partition on
appetitive source count if you need the two halves separately. Every explore
feed_settle comes from the sourced half - there are zero settles without a source
(verified).
aversive_avoid is a reflex probe, not a competence test
105/105 episodes end left_arena; path straightness 0.92 vs 0.61 corpus-wide;
median final displacement 150 mm; median minimum distance to the aversive source 28.5
mm against an 8 mm threshold. Its 100% "success" cannot fail by construction - read
it as a reflex sanity check, not a solved task.
Fall mechanism
28/28 fall episodes are preceded by an obstacle contact (82% within 5 windows, 100%
within 10 and 20 windows, vs 3.2% / 6.3% / 12.2% by chance). The model has no
righting reflex - there is no proprioceptive loop - so an inverted fly cannot
recover; the episode ends after 20 windows inverted (end_reason='fell').
Schema
schema_version embodied-0.5 / embodied-0.6. label/outcomes columns:
| col | name | meaning |
|---|---|---|
| 0 | dist | distance to the nearest appetitive source (mm) |
| 1 | reached | within FEED_RADIUS of an appetitive source |
| 2 | collision | MuJoCo CONTACT-based obstacle collision (not a disc approximation) |
| 3 | fell | inverted / airborne |
| 4 | aversive_dist | distance to the nearest aversive source (nan if none) |
end_reason values: feed_settle (settled at food), left_arena (departed
150 mm from spawn),
completed(ran to the window cap),fell(inverted, no recovery),physics_error(MuJoCo instability, episode truncated). Splits: train / val / test.
Provenance: the corpus legitimately carries TWO git SHAs - 745cd1a (69 untouched
shards) and 61e107d (12 re-run cells: strong_vs_weak x9 + 3 smoke top-ups). Their
CONTROL CONSTANTS are identical, so they are compatible. corpus_version = v2,
patched in place from an initial mislabel 'v1' with corpus_version_patched_from
recorded on each affected shard (not hidden).
The two SHAs also explain the mixed schema above: the 12 re-run shards are
embodied-0.6 and additionally carry meta/vision_mode, because they were
collected AFTER the vision-provenance commit landed; the 69 untouched shards are
embodied-0.5. This is history, not sloppiness - both schemas are
backward-compatible and every reader handles either. All shards ran retinotopic
vision (the retina fix predates both SHAs).
Corpus statistics
The block below is emitted verbatim by corpus_report so the card cannot drift from
the data:
corpus: 81 shards, 945 episodes, 142689 windows, 44.78 GB on disk
bytes/window = 314 KB (v1 was 761 KB/window)
success is PER-FAMILY (not comparable across families): approach=reached (<7mm / feed_settle); feed=reached (<7mm / feed_settle); avoid=min(aversive_dist) > 8mm; escape=escape label while loom > 0.5; forage/explore=n/a
scenario diff split eps len med/mean succ% reach% coll% fell% H(act) family end_reasons
aversive_avoid easy test 10 185/ 186 100.0% 0.0% 0.0% 0.0% 1.45 avoid:10 left_arena:10
aversive_avoid easy train 20 180/ 184 100.0% 0.0% 0.0% 0.0% 1.44 avoid:20 left_arena:20
aversive_avoid easy val 5 187/ 186 100.0% 0.0% 0.0% 0.0% 1.36 avoid:5 left_arena:5
aversive_avoid hard test 10 182/ 183 100.0% 0.0% 0.0% 0.0% 1.48 avoid:10 left_arena:10
aversive_avoid hard train 20 182/ 184 100.0% 0.0% 0.0% 0.0% 1.45 avoid:20 left_arena:20
aversive_avoid hard val 5 182/ 182 100.0% 0.0% 0.0% 0.0% 1.49 avoid:5 left_arena:5
aversive_avoid medium test 10 186/ 187 100.0% 0.0% 0.0% 0.0% 1.40 avoid:10 left_arena:10
aversive_avoid medium train 20 182/ 181 100.0% 0.0% 0.0% 0.0% 1.48 avoid:20 left_arena:20
aversive_avoid medium val 5 183/ 182 100.0% 0.0% 0.0% 0.0% 1.46 avoid:5 left_arena:5
explore easy test 10 76/ 92 100.0% 80.0% 0.0% 0.0% 1.72 approach:8, explore:2 settled:8, left_arena:2
explore easy train 20 168/ 138 100.0% 40.0% 0.0% 0.0% 1.46 explore:12, approach:8 left_arena:11, settled:8, timeout:1
explore easy val 5 172/ 182 n/a 0.0% 0.0% 0.0% 1.26 explore:5 left_arena:4, timeout:1
explore hard test 10 154/ 140 100.0% 50.0% 0.0% 0.0% 1.58 approach:5, explore:5 settled:5, left_arena:5
explore hard train 20 166/ 218 100.0% 30.0% 0.0% 0.0% 1.40 explore:14, approach:6 left_arena:8, settled:6, timeout:6
explore hard val 5 84/ 116 100.0% 60.0% 0.0% 0.0% 1.36 approach:3, explore:2 settled:3, left_arena:2
explore medium test 10 131/ 158 100.0% 50.0% 0.0% 0.0% 1.51 approach:5, explore:5 settled:5, left_arena:4, timeout:1
explore medium train 20 83/ 111 100.0% 65.0% 0.0% 0.0% 1.65 approach:13, explore:7 settled:13, left_arena:7
explore medium val 5 165/ 139 100.0% 40.0% 0.0% 0.0% 1.53 explore:3, approach:2 left_arena:2, settled:2, timeout:1
gauntlet easy test 10 168/ 185 70.0% 70.0% 20.0% 0.0% 1.86 approach:10 settled:7, left_arena:3
gauntlet easy train 20 125/ 140 75.0% 75.0% 15.0% 5.0% 1.85 approach:20 settled:15, left_arena:4, fell:1
gauntlet easy val 5 177/ 170 60.0% 60.0% 0.0% 0.0% 1.75 approach:5 settled:3, left_arena:2
gauntlet hard test 10 176/ 184 10.0% 10.0% 0.0% 0.0% 1.46 approach:10 left_arena:9, settled:1
gauntlet hard train 20 186/ 208 15.0% 15.0% 0.0% 0.0% 1.44 approach:20 left_arena:15, settled:3, timeout:2
gauntlet hard val 5 174/ 162 20.0% 20.0% 0.0% 0.0% 1.55 approach:5 left_arena:4, settled:1
gauntlet medium test 10 194/ 208 20.0% 20.0% 20.0% 0.0% 1.69 approach:10 left_arena:7, settled:2, timeout:1
gauntlet medium train 20 190/ 191 35.0% 35.0% 25.0% 5.0% 1.61 approach:20 left_arena:10, settled:7, timeout:2, fell:1
gauntlet medium val 5 177/ 183 40.0% 40.0% 0.0% 0.0% 1.60 approach:5 left_arena:2, settled:2, timeout:1
good_vs_bad easy test 10 58/ 62 100.0% 100.0% 0.0% 0.0% 1.75 approach:10 settled:10
good_vs_bad easy train 20 58/ 61 100.0% 100.0% 0.0% 0.0% 1.77 approach:20 settled:20
good_vs_bad easy val 5 60/ 60 100.0% 100.0% 0.0% 0.0% 1.74 approach:5 settled:5
good_vs_bad hard test 10 124/ 122 100.0% 100.0% 0.0% 0.0% 1.66 approach:10 settled:10
good_vs_bad hard train 20 110/ 114 100.0% 100.0% 0.0% 0.0% 1.75 approach:20 settled:20
good_vs_bad hard val 5 106/ 120 80.0% 80.0% 0.0% 0.0% 1.68 approach:5 settled:4, left_arena:1
good_vs_bad medium test 10 90/ 116 100.0% 100.0% 0.0% 0.0% 1.79 approach:10 settled:10
good_vs_bad medium train 20 92/ 118 95.0% 95.0% 0.0% 0.0% 1.86 approach:20 settled:19, timeout:1
good_vs_bad medium val 5 93/ 100 100.0% 100.0% 0.0% 0.0% 1.77 approach:5 settled:5
meadow easy test 10 295/ 271 100.0% 100.0% 30.0% 20.0% 1.86 approach:8, escape:2 timeout:8, fell:2
meadow easy train 20 346/ 353 100.0% 100.0% 15.0% 5.0% 1.92 approach:17, escape:2, feed:1 timeout:17, left_arena:1, settled:1, fell:1
meadow easy val 5 200/ 264 100.0% 100.0% 40.0% 0.0% 2.04 approach:4, escape:1 timeout:3, settled:1, left_arena:1
meadow hard test 10 200/ 219 90.0% 90.0% 70.0% 10.0% 2.02 escape:5, approach:5 settled:5, timeout:3, fell:1, left_arena:1
meadow hard train 20 246/ 270 90.0% 90.0% 50.0% 5.0% 2.08 approach:14, escape:6 timeout:9, left_arena:6, settled:4, fell:1
meadow hard val 5 261/ 239 100.0% 100.0% 80.0% 20.0% 1.92 escape:3, approach:2 timeout:2, settled:2, fell:1
meadow medium test 10 298/ 269 90.0% 80.0% 60.0% 40.0% 2.16 escape:6, approach:4 fell:4, timeout:3, left_arena:2, settled:1
meadow medium train 20 240/ 260 95.0% 95.0% 60.0% 15.0% 2.04 approach:14, escape:6 timeout:10, fell:3, settled:3, left_arena:3, physics:1
meadow medium val 5 378/ 339 100.0% 100.0% 60.0% 0.0% 1.92 approach:5 timeout:4, left_arena:1
obstacle_field easy test 10 86/ 108 80.0% 80.0% 30.0% 10.0% 1.89 approach:10 settled:8, fell:1, left_arena:1
obstacle_field easy train 20 82/ 93 90.0% 90.0% 30.0% 10.0% 1.88 approach:20 settled:17, fell:2, timeout:1
obstacle_field easy val 5 82/ 66 60.0% 60.0% 60.0% 20.0% 1.70 approach:5 settled:3, physics:1, fell:1
obstacle_field hard test 10 92/ 154 80.0% 80.0% 40.0% 20.0% 1.86 approach:10 settled:5, fell:2, timeout:2, left_arena:1
obstacle_field hard train 20 87/ 161 65.0% 65.0% 60.0% 15.0% 1.78 approach:20 settled:11, left_arena:3, timeout:3, fell:2, physics:1
obstacle_field hard val 5 81/ 100 80.0% 80.0% 20.0% 0.0% 1.88 approach:5 settled:4, left_arena:1
obstacle_field medium test 10 116/ 124 70.0% 70.0% 10.0% 0.0% 1.85 approach:10 settled:7, left_arena:3
obstacle_field medium train 20 130/ 141 60.0% 60.0% 55.0% 20.0% 1.75 approach:20 settled:12, fell:3, left_arena:3, timeout:2
obstacle_field medium val 5 95/ 112 80.0% 80.0% 40.0% 0.0% 1.83 approach:5 settled:4, left_arena:1
scent_trail easy test 10 70/ 82 100.0% 100.0% 0.0% 0.0% 1.76 feed:9, approach:1 settled:10
scent_trail easy train 20 99/ 107 100.0% 100.0% 0.0% 0.0% 1.75 feed:19, approach:1 settled:20
scent_trail easy val 5 85/ 83 100.0% 100.0% 0.0% 0.0% 1.77 feed:4, approach:1 settled:5
scent_trail hard test 10 190/ 221 100.0% 100.0% 0.0% 0.0% 1.25 approach:10 timeout:5, settled:5
scent_trail hard train 20 314/ 274 100.0% 100.0% 0.0% 0.0% 1.15 approach:20 timeout:13, settled:7
scent_trail hard val 5 386/ 357 100.0% 100.0% 0.0% 0.0% 1.30 approach:5 timeout:5
scent_trail medium test 10 202/ 227 100.0% 100.0% 0.0% 0.0% 1.38 approach:10 timeout:7, settled:3
scent_trail medium train 20 212/ 227 100.0% 100.0% 0.0% 0.0% 1.44 approach:19, feed:1 timeout:10, settled:10
scent_trail medium val 5 76/ 142 100.0% 100.0% 0.0% 0.0% 1.74 approach:5 settled:4, timeout:1
strong_vs_weak easy test 10 53/ 54 100.0% 100.0% 0.0% 0.0% 1.81 approach:10 settled:10
strong_vs_weak easy train 20 54/ 55 100.0% 100.0% 0.0% 0.0% 1.92 approach:20 settled:20
strong_vs_weak easy val 5 60/ 69 100.0% 100.0% 0.0% 0.0% 1.80 approach:5 settled:5
strong_vs_weak hard test 10 94/ 98 100.0% 100.0% 0.0% 0.0% 1.86 approach:10 settled:10
strong_vs_weak hard train 20 96/ 100 100.0% 100.0% 0.0% 0.0% 1.86 approach:20 settled:20
strong_vs_weak hard val 5 88/ 94 100.0% 100.0% 0.0% 0.0% 1.91 approach:5 settled:5
strong_vs_weak medium test 10 74/ 77 100.0% 100.0% 0.0% 0.0% 1.89 approach:10 settled:10
strong_vs_weak medium train 20 69/ 74 100.0% 100.0% 0.0% 0.0% 1.86 approach:20 settled:20
strong_vs_weak medium val 5 71/ 75 100.0% 100.0% 0.0% 0.0% 1.67 approach:5 settled:5
sugar_approach easy test 10 56/ 57 100.0% 100.0% 0.0% 0.0% 1.66 approach:10 settled:10
sugar_approach easy train 20 56/ 56 100.0% 100.0% 0.0% 0.0% 1.69 approach:20 settled:20
sugar_approach easy val 5 58/ 57 100.0% 100.0% 0.0% 0.0% 1.70 approach:5 settled:5
sugar_approach hard test 10 88/ 88 100.0% 100.0% 0.0% 0.0% 1.68 approach:10 settled:10
sugar_approach hard train 20 86/ 88 100.0% 100.0% 0.0% 0.0% 1.73 approach:20 settled:20
sugar_approach hard val 5 89/ 88 100.0% 100.0% 0.0% 0.0% 1.73 approach:5 settled:5
sugar_approach medium test 10 69/ 74 100.0% 100.0% 0.0% 0.0% 1.81 approach:10 settled:10
sugar_approach medium train 20 71/ 71 100.0% 100.0% 0.0% 0.0% 1.85 approach:20 settled:20
sugar_approach medium val 5 64/ 67 100.0% 100.0% 0.0% 0.0% 1.74 approach:5 settled:5
TIME-TO-REACH: median episode length (windows) by scenario x difficulty
scenario easy medium hard monotone(e<m<h)
aversive_avoid 184 183 182 NO
explore 161 88 166 NO
gauntlet 161 191 178 NO
good_vs_bad 58 92 112 yes
meadow 322 265 241 NO
obstacle_field 82 120 86 NO
scent_trail 91 200 305 yes
strong_vs_weak 54 72 93 yes
sugar_approach 56 69 87 yes
POST-FALL WINDOWS (degenerate tail FIX D removes):
corpus: 1501/142689 = 1.05% (v1 was 6.0%)
gauntlet 40/19122 = 0.21%
meadow 434/29462 = 1.47%
obstacle_field 1027/13142 = 7.81%
COLLISION: contact path vs the old 0.6mm disc (FIX F quantified)
scenario contact% disc% ratio windows
gauntlet 0.61% 0.02% 29.25 19122
meadow 3.28% 0.08% 38.68 29462
obstacle_field 3.43% 0.07% 50.11 13142
corpus contact=1.08% disc=0.03% ratio=40.39 (>1 means contacts caught what the disc missed)
MI(action; scenario) = 0.197 bits (over 142689 windows)
MI(action; family) = 0.332 bits (family = per-window `expected` task label; over 142689 windows)
PROVENANCE: multi-SHA corpus - control constants AGREE across all shards, so it is COMPATIBLE (exit 0). SHA inventory:
git_sha=61e107dffd81: 12 shard(s)
git_sha=745cd1a2174a: 69 shard(s)
diff between SHAs: only git_sha (+ any corpus_version); the control constants below are identical:
pool_yaw_gain=-2 pool_yaw_max=3 taxis_sign=1 ema_alpha=0.45 drive_min=0.2 drive_max=1.4 feed_arrest_odor=0.15 feed_arrest_span=0.3
Limitations
- aversive_avoid and explore are CONTROLS, not solvable tasks (see above); do not fold their success into a headline number.
- 3
physics_errorstubs in the corpus (one only 25 windows long). - obstacle_field still has 7.81% post-fall windows.
- MI(action;scenario) = 0.197 is low PARTLY BY DESIGN: several scenarios share one policy ("walk up the odour gradient"), so the action need not distinguish them. Read MI(action;family) alongside it.
- No righting reflex (see the fall mechanism above).
Licence
TODO(maintainer): choose and set a licence. Deliberately not invented here.
- Downloads last month
- 9