ChronoLoop checkpoints

ChronoLoop adds a small recurrent memory to LoopWAM v0: 16 memory tokens ride through the pre-blocks and the shared core loops with the observation, are read out after the last core loop, squashed and mixed into a carried state, and are fed back at the next replanning query. Non-memory tokens read the memory through a per-head gate tanh(alpha) that starts at 0, so every run starts exactly at its parent.

Every run fine-tunes the same parent, anhdao69/LoopWAM_NT/libero-all-suites/v0-video4-action4 (LoopWAM v0, 4 video / 4 action core loops, 388/400 on the four LIBERO suites), on all four LIBERO suites (1,712 demonstrations, 277,713 windows) for 10 epochs with global batch 128 (32 streams x 4 consecutive windows, truncated BPTT S = 4). The schedule, noise and data order are shared by all runs, so they differ only in the memory flags.

Folder	Run	Definition
CL-0_no-memory_a4	CL-0	No memory (control), K_a = 4
CL-A_mem16-learned-loopwrite_a4	CL-A	Memory written by the shared core loops (default ChronoLoop)
CL-REG_mem16-reset-registers_a4	CL-REG	Same 16 tokens reset to s = 0 every query
CL-W2_mem16-learned-external-updater_a4	CL-W2	Read-only memory in the backbone; separate 1-block updater
CL-0-at1_no-memory_a1	CL-0@1	CL-0 with 1 action core loop
CL-A-at1_mem16-learned-loopwrite_a1	CL-A@1	CL-A with 1 action core loop
CL-FRAME_history-frame-k3_a4	CL-FRAME	No memory; observation from query k-3 as extra gated clean frame

Each folder has epoch_08/, epoch_09/ and epoch_10/ with policy.pt (weights only, format chronoloop-v1) and config.json, plus the run's manifest.json, metrics.jsonl and timing.json. A folder appears only after its training has written those epochs, and every upload is checked against the local SHA-256.

Load with the chrono branch of LoopWAM_NT:

from fastwam.models.wan22.chronoloop import ChronoConfig, create_chronoloop
import torch
payload = torch.load('policy.pt', map_location='cpu', weights_only=False, mmap=True)
model, _ = create_chronoloop(ChronoConfig(**payload['chrono']), 'Wan2.1_VAE.pth', checkpoint_path='policy.pt', device='cuda')
# per replanning query: out = model.chrono_infer_action(image, proprio, context, mask, state=state); state = out['state']

Rollouts: scripts/evaluate_chronoloop_libero.py (700 policy steps, 30 settling steps, replan every 10 actions, memory reset at episode start, fixed or fresh noise).