ACRFT YAM patch-critics (s347)

Frozen-DINOv2 patch critics trained on the YAM lego-taxi teleop set (347 episodes: 300 success / 47 failure), for critic-guided rollouts (best-of-N / adaptive-chunk selection) on top of a pi0.5 BC policy. All arms share the house value conventions: cost_to_goal reward (-1/step), gamma 0.99964, support [-2777.8, 0] (HL-Gauss 101 atoms), h_goal=30 absorbing goal, homing tails truncated from every episode, K=2 ensemble, 200k steps, batch 256, lr 1e-4.

Input contract (per config.json input_spec): 3 cameras (agentview + both wrists) at 224, pi0.5-normalized (quantile) state, pi0.5-normalized JOINT-DELTA action chunks (horizon 30), proprio = joint positions + grippers (14) unless noted. The serving wrapper reproduces this preprocessing from the spec automatically.

Arms

folder chunking tau q_reduction feat aug note
patch_critic_yam_s347_fixed_200k fixed (single prefix H=30) 0.7 min none DEAS-style fixed-chunk IQL baseline
patch_critic_yam_s347_g5_200k adaptive (prefixes 5..30) 0.7 min none adaptive-chunk IQL (ARQ)
patch_critic_yam_s347_g5_tau9_min adaptive 0.9 min 0.1/0.05 hindsight-leakage-reduced arm
patch_critic_yam_s347_g5_tau9_mean adaptive 0.9 mean 0.1/0.05 leakage-reduced, mean ensemble
patch_critic_yam_s347_g5all_200k adaptive 0.7 min none proprio ALL(42) incl. velocity/effort - extra sensors, NOT baseline-fair

Dataset-side evaluation (stride-20, 347 eps, scorer 673f33a)

fixed_200k g5_200k
success V(s0) -1620.7 -1620.4
failure V(s0) -2580.4 -2606.8
success Vslope (vs true cost-to-go) 0.85 0.90
success V at goal-1 -298.5 -36.6
k* success / failure 30 / 30 (trivial) 12.8 / 6.1

Known caveat: the tau=0.7 no-aug arms carry strong hindsight leakage (success/failure V(s0) gap ~1000 where the behaviour-policy ideal is 0); the tau9 arms reduce that gap by ~65%. Compare arms relatively; do not read absolute values as calibrated returns.

Rollout / serving

uv run python scripts/serve_patch_critic.py \
  --config pi05_yam_lego_taxi --ckpt <pi05 BC checkpoint> \
  --critic-dir <downloaded critic folder> \
  --num-samples 8 --mode adaptive     # or --mode bon

(alpha-Flow policy: add --config pi05_yam_lego_taxi_alphaflow --flow-steps 1 for one-step sampling; see jellyho/pi05_yam_lego_taxi_alphaflow.)

Trained by the ACRFT patch-critic session (code: jellyho/ACRFT integration branch, scripts/train_patch_critic_cached.py; feature cache yam_s347). An mc_floor-off pair (fixed/g5) is training and will be added when complete.

Downloads last month

-

Downloads are not tracked for this model. How to track
Video Preview
loading