screw_t3tiny_ingraph_50k: T3-tiny fine-tuned in-graph ACT (best val 0.2737 @ 4600) + TS encoder + code (t3 branch, dim-agnostic feature_policy)

screw_sitr5_ingraph_50k FINAL: best val 0.2603 @ step 8200 (early stop 9200); refreshed ckpt + TS encoder + doc

screw_sitr5_ingraph_50k: in-graph fine-tuned SITR ACT (snapshot step 7200) + TS encoder export + code (sitr branch, bf16 flag, feature_policy in-graph loading)

Qistee/HTT integration: feature_policy.py + README (validated bit-identical)

TorchScript SITR feature encoder

TorchScript T3 feature encoder (self-contained, validated)

Add training recipe doc

Add train.py — training now self-contained with the episodes dataset repo

Add feature-preprocessing doc

HTT tactile_checkpoint deploy fixes: bg capture at episode-start pose, n_action_steps=30 restored, 30 Hz control

Add fixed robot runner for T3/SITR + deploy findings/fixes doc

Add all intermediate + final(50K) checkpoints for both runs + CHECKPOINT_MAP

Add remote-workstation encoder setup doc

Add encoder weights (t3_medium, SITR_B18) + self-test episode

Reference inference stack + validated live_features.py + UPDATE_GUIDE for target-machine merge

Add deploy doc under its own filename (duplicate of README)

T3 + SITR ACT screw baselines (act_best, args, encode tools, deploy doc)

initial commit