craigm26/kith-valley-walks

Kith: walks through the valley

Episodes from Kith, a wordless browser game in which a person and the small robot ducks by Pollen Robotics (Microduck) walk somewhere together. Here it is the walk through the valley: carrying the last seed of the great tree home, with the windmill, the well, the tangles, the windy ridge, the dark wood, the fire and the water tower, and the grove. Positions are at the game's scale; the routine in each walk replays it at the robot's pace.

Each episode is one Microduck's side of one walk: what it could sense, five times a second, and what it did next, with the game's reward and what it was doing, in words. The point is a policy that keeps company the way people do: at someone's pace, waiting, answering, asking for help and giving it, and doing its part.

This release is synthetic. No one has given a walk yet, so every episode here is the game played by simple scripted players (and the game's own ducks). It shows the format end to end and gives a baseline. The episodes people give from the end of a walk at kithgame.com go out in the next releases, marked gift.

Version 0.1.0. 9 episodes, 10,086 frames (0.56 hours) at 5 fps. By source: synthetic 9. Whose duck: the game 6, a scripted player 3.

A frame

• observation.state (35): speed, face, sitting, fallen, skating, held, trav_seen, trav_ahead_m, trav_dist_m, trav_moving, trav_song, near_seen, near_ahead_m, near_dist_m, near_game, near_human, near_agent, others_2m, heard, heard_ahead, wind, ash, slope, chapter, need_seen, need_ahead_m, need_wake, need_bite, need_climb, need_lamp, need_dig, need_stone, need_seed, need_ask, need_wait. Robot-relative: "ahead" is the way the duck faces, distances are metres at the robot's scale, speed is a fraction of its top speed (0 while it's carried: held). *_seen says whether the value next to it is known (0 means it is not). near_* is the nearest one other than the traveler. need_* is what the way needed from a duck right then, nearest first.
• action (27): walk, walk_speed, sit, quack, voice_chirp, voice_greet, voice_inquire, voice_alarm, voice_coo, voice_wheee, voice_peck, act_hop, act_roll, act_kick, act_bow, act_nod, act_shake, act_peck, act_kick_left, act_flamingo, act_glance, act_climb, act_bite, gaze_ahead, gaze_up_at_traveler, gaze_at_singer, gaze_back. What it did until the next frame. walk is the control held (-1, 0, 1), walk_speed how fast as a fraction of top speed, quack how strongly and voice_* what it meant (the real Microduck's voice bank), act_* a move, gaze_* where it looked. These drive the robot's own skills (the walk policy's velocity command, its trained moves), not its joints.
• next.reward: the published schedule (rewards version 2): a little for every frame kept close to the traveler, more for the moments reached together, less for what was missed.
• next.done: the walk's last frame.
• task: what the duck was doing, in words, from the game's own labelling.

The moments in these episodes: answered_song 636, led_the_way 97, walked_together 52, crossed_ridge_together 48, waited 20, fetched 19, cleared_the_way 18, gathered_at_tree 9, woke_the_traveler 3, freed_the_mill 3, opened_the_well 3, lit_the_way 3, put_out_the_fire 3, dug_for_the_seed 3.

Watch a walk

Every episode here plays again in the game's own scene, with a timeline of what the duck was doing, its reward and its moments, and its row of observations and actions at the cursor: kithgame.com/replay?ds=craigm26/kith-valley-walks lists them.

Everything else, in kith/

• kith/views.jsonl: one line per episode (in episode_index order): where it came from, who walked the duck and the traveler (human, script, game), who else was there, its scorecard, and its moments.
• kith/episodes.jsonl.gz: each episode as the game exports it (kith.episodes/3): every participant's frames, the songs, moves and moments, the feedback, the misses, the segments, the scorecard, the routine as steps for a real Microduck (microduck_rl task ids and commands), and an OpenCastor plan.

Where it comes from, and privacy

Walks are given from the end card of a walk at kithgame.com, by the person who walked it. A walk taken back leaves the next release (one already out keeps it). They hold game state only: positions, the control held, poses, songs and quacks and what they meant, moves, game moments, pats and nods. Everyone is renamed (self, p1, p2...), there is no text and nothing from the network, and people who chose "Don't share my play" are never in them. An episode is released for the giver's own duck and for the ducks the game walked; another person's or an agent's duck appears only as someone nearby. Seed walks are played by scripted players and say so (source: synthetic, walked_by: script).

Licence and credits

CC BY 4.0. Made by Craig Merry and Claude. The Microduck is by Pollen Robotics.

from lerobot.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("craigm26/kith-valley-walks")
frame = ds[0]   # observation.state, action, next.reward, task...

meta/info.json

{
  "codebase_version": "v3.0",
  "fps": 5,
  "features": {
    "observation.state": {
      "dtype": "float32",
      "shape": [
        35
      ],
      "names": [
        "speed",
        "face",
        "sitting",
        "fallen",
        "skating",
        "held",
        "trav_seen",
        "trav_ahead_m",
        "trav_dist_m",
        "trav_moving",
        "trav_song",
        "near_seen",
        "near_ahead_m",
        "near_dist_m",
        "near_game",
        "near_human",
        "near_agent",
        "others_2m",
        "heard",
        "heard_ahead",
        "wind",
        "ash",
        "slope",
        "chapter",
        "need_seen",
        "need_ahead_m",
        "need_wake",
        "need_bite",
        "need_climb",
        "need_lamp",
        "need_dig",
        "need_stone",
        "need_seed",
        "need_ask",
        "need_wait"
      ]
    },
    "action": {
      "dtype": "float32",
      "shape": [
        27
      ],
      "names": [
        "walk",
        "walk_speed",
        "sit",
        "quack",
        "voice_chirp",
        "voice_greet",
        "voice_inquire",
        "voice_alarm",
        "voice_coo",
        "voice_wheee",
        "voice_peck",
        "act_hop",
        "act_roll",
        "act_kick",
        "act_bow",
        "act_nod",
        "act_shake",
        "act_peck",
        "act_kick_left",
        "act_flamingo",
        "act_glance",
        "act_climb",
        "act_bite",
        "gaze_ahead",
        "gaze_up_at_traveler",
        "gaze_at_singer",
        "gaze_back"
      ]
    },
    "next.reward": {
      "dtype": "float32",
      "shape": [
        1
      ],
      "names": null
    },
    "next.done": {
      "dtype": "bool",
      "shape": [
        1
      ],
      "names": null
    },
    "timestamp": {
      "dtype": "float32",
      "shape": [
        1
      ],
      "names": null
    },
    "frame_index": {
      "dtype": "int64",
      "shape": [
        1
      ],
      "names": null
    },
    "episode_index": {
      "dtype": "int64",
      "shape": [
        1
      ],
      "names": null
    },
    "index": {
      "dtype": "int64",
      "shape": [
        1
      ],
      "names": null
    },
    "task_index": {
      "dtype": "int64",
      "shape": [
        1
      ],
      "names": null
    }
  },
  "total_episodes": 9,
  "total_frames": 10086,
  "total_tasks": 28,
  "chunks_size": 1000,
  "data_files_size_in_mb": 100,
  "video_files_size_in_mb": 200,
  "data_path": "data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet",
  "video_path": null,
  "robot_type": "microduck",
  "splits": {
    "train": "0:9"
  }
}