Ali-Uraish commited on
Commit
7dfe6bc
·
verified ·
1 Parent(s): cbf35a0

Add files using upload-large-folder tool

Browse files
Files changed (1) hide show
  1. chess-sim/README.md +11 -0
chess-sim/README.md CHANGED
@@ -8,6 +8,7 @@ Every dataset is LeRobot v3.0 at 30 fps, 640x480 images, with the source square
8
 
9
  | folder | episodes | frames | cameras | video | notes |
10
  |---|---|---|---|---|---|
 
11
  | `datasets/baseline_v1/` | 300 | 89008 | overhead, wrist | h264 yuv444p | The simple baseline (see *Baseline scene*): one fixed scene and ten fixed opening moves. Trains `Machanize/chess_phase_smolvla_baseline`. |
12
  | `datasets/pilot_60/` | 60 | 17829 | overhead, rover, wrist | h264 yuv420p | The first pilot: nominal board and tray, narrower lighting and camera variety. Kept for reference. |
13
  | `datasets/varied_2000/` | 2000 | 594702 | overhead, rover, wrist | h264 yuv420p | Heavy randomisation (see *What varies*). Trained `Machanize/chess_phase_smolvla`. |
@@ -36,6 +37,16 @@ Everything fixed, to get reliable closed-loop success before adding variety back
36
  - Markers with an opaque 2 px outline; pieces centred in their squares.
37
  - Sharper video: H.264 at CRF 18 with full 4:4:4 chroma.
38
 
 
 
 
 
 
 
 
 
 
 
39
  Expert validation (random layouts): 99.02% of 1536 trials (target above 99.0%).
40
 
41
  ## Load
 
8
 
9
  | folder | episodes | frames | cameras | video | notes |
10
  |---|---|---|---|---|---|
11
+ | `datasets/baseline_recovery_v1/` | 120 | 36645 | overhead, wrist | h264 yuv444p | Correction demonstrations in the baseline scene (see *Corrections*): the baseline policy drives, the teacher takes over. With baseline_v1, trains `Machanize/chess_phase_smolvla_baseline_r2`. |
12
  | `datasets/baseline_v1/` | 300 | 89008 | overhead, wrist | h264 yuv444p | The simple baseline (see *Baseline scene*): one fixed scene and ten fixed opening moves. Trains `Machanize/chess_phase_smolvla_baseline`. |
13
  | `datasets/pilot_60/` | 60 | 17829 | overhead, rover, wrist | h264 yuv420p | The first pilot: nominal board and tray, narrower lighting and camera variety. Kept for reference. |
14
  | `datasets/varied_2000/` | 2000 | 594702 | overhead, rover, wrist | h264 yuv420p | Heavy randomisation (see *What varies*). Trained `Machanize/chess_phase_smolvla`. |
 
37
  - Markers with an opaque 2 px outline; pieces centred in their squares.
38
  - Sharper video: H.264 at CRF 18 with full 4:4:4 chroma.
39
 
40
+ ## Corrections (baseline_recovery_v1)
41
+
42
+ DAgger-style data in the baseline scene. The baseline policy (`Machanize/chess_phase_smolvla_baseline`) drives; the scripted teacher takes over from exactly where the arm is, and only the teacher's part is recorded. A grasp inspection had shown the policy's jaws close 4-6 mm off the piece when it fails, against about 1 mm when it holds. Handovers (`recovery.trigger` in `phase2_episodes.jsonl`):
43
+
44
+ - `pre_grasp`: the jaws, still open, come within 15-45 mm above the piece. The teacher lines up above it, holds still 0.5 s (the wrist camera sees the piece between the jaws), descends, grasps and finishes the move.
45
+ - `misaligned_close`: the policy starts closing more than 2.5 mm off. The teacher reopens, rises clear, holds still 0.5 s to look, lines up, grasps and finishes.
46
+ - `failed_lift`: the jaws closed but the piece did not rise within 1.2 s. Same recovery.
47
+
48
+ Episodes where the piece had toppled or a neighbour had moved before the handover are dropped, as are failed recoveries.
49
+
50
  Expert validation (random layouts): 99.02% of 1536 trials (target above 99.0%).
51
 
52
  ## Load