teawhite/hollow-knight-boss-fights

Hollow Knight Boss Fights — autoheal, frame-aligned video + state

Autonomous-bot boss-fight recordings from Hollow Knight (classic 1.5.78), captured for imitation learning / world-model / decision-transformer research. Each session pairs a screen-recorded video with a frame-aligned state log (player + boss telemetry at ~20 Hz) and an event-driven keyboard-input log. A bot fights each boss autonomously; the player takes real damage (hit reactions, i-frames, and a visible HP sawtooth) but is kept alive by a mod hook ("autoheal" mode), so trajectories are long and continuous without deaths.

• 114 sessions · 33.7 hours · ~57 GB
• Modality: RGB gameplay video (.mp4) + JSON-lines state + JSON-lines input
• Alignment: every state/input record carries video_t (SMPTE HH:MM:SS:FF) and videoSeconds, so records map directly onto video frames.

Bosses × arenas

Bosses are placed in arenas via cross-arena injection to decorrelate boss identity from background: two bosses are injected into the Broken Vessel arena so a model can't shortcut "arena ⇒ boss".

Group (folder)	Boss	Arena	Placement	Sessions	Hours
broken_vessel__GG_Broken_Vessel	Broken Vessel	Ancient Basin (GG_Broken_Vessel)	native	48	11.37
hornet_1__GG_Broken_Vessel	Hornet	Ancient Basin (GG_Broken_Vessel)	injected (cross-arena)	31	11.06
gruz_mother__GG_Broken_Vessel	Gruz Mother	Ancient Basin (GG_Broken_Vessel)	injected (cross-arena)	21	6.12
dung_defender__GG_Dung_Defender	Dung Defender	Royal Waterways (GG_Dung_Defender)	native	14	5.12

Directory layout

data/<group>/<session_id>/
    video.mp4       # screen recording of the fight (OBS)
    state.jsonl     # ~20 Hz player + boss + game state, frame-aligned
    input.jsonl     # keyboard down/up events, frame-aligned
    session.json    # session id, OBS start/stop UTC, duration, boss/arena/mode
metadata.jsonl      # one row per session (id, group, boss, arena, mode, duration, file paths)
SCHEMA.md           # full field reference for state.jsonl / input.jsonl

Collection details

• Mode: autoheal — the bot takes real hits (recoil, flash, i-frames, logged HP drop) but a mod hook caps otherwise-lethal damage and refills when low, so HP sawtooths (typically 3–5) and the player never dies.
• Nail damage: 2 (low, for long fights).
• Boss placement: the Godhome single-boss arenas; injected bosses are spawned into a foreign arena.
• Video↔state alignment: OBS recording start is timestamped; each record's video_t/videoSeconds are computed against that origin. Use them to index frames (don't assume a fixed offset).

Example: load a session

import json, cv2
base = "data/hornet_1__GG_Broken_Vessel/<session_id>"
states = [json.loads(l) for l in open(f"{base}/state.jsonl", encoding="utf-8")]
cap = cv2.VideoCapture(f"{base}/video.mp4")
s = states[100]
cap.set(cv2.CAP_PROP_POS_MSEC, s["videoSeconds"] * 1000)  # align state -> frame
ok, frame = cap.read()
print(s["boss"]["name"], s["boss"]["action"], s["player"]["hp"])

Intended uses & limitations

• Uses: imitation learning, behavior cloning, world models, video+action modeling, boss-move recognition.
• Not human play: actions come from a scripted/autonomous bot, not expert humans; treat as bot-policy data.
• autoheal bias: the player rarely dies, so the data under-represents failure/death states.
• Alignment caveat: the first ~1–2 frames of a session may show the arena's native boss before an injected boss takes over; trim the first moment if needed.

Provenance & IP

Recorded from Hollow Knight by Team Cherry using a custom Modding-API mod. The dataset is released for research; game visuals/audio remain the property of Team Cherry. session.json has been sanitized (local paths / OS username removed).

License

Dataset annotations/logs: CC-BY-4.0. Underlying game footage is Team Cherry's copyrighted material, included here for research use.