Yesterday
-
root README: press_button_openvla_delta control row
-
press_button_terra_delta README: matched control result
-
press_button OpenVLA control (delta): 1.92 deg vs TERRA 0.92, ignores the instruction
-
press_button_terra_delta README: ensembling undecided for delta
-
README: add press_button_terra_delta row
-
press_button_terra_delta: TERRA arm, delta actions, 3000 steps (README: delta interface, held-out + lang-swap)
2 days ago
-
Handoff: the base is public now, no token needed
-
Card: document p2fft_20k, the phase-2 base, and that action_decoder.pt is required
-
Phase-2 pretrained base (OpenVLA-7B, 20k full-parameter steps, latp2_inv latent) as a post-training INIT
-
Handoff: clone URL, the settled padding answer, confirmed task names and episode lengths
-
Card: the real code URL (owner was missing, sent a reader to a 404) and drop the wrong `private: true`
-
Handoff brief: training the press_button arm on a Jetstream2 g3.xl
4 days ago
-
piper README: null baselines -- the policy beats the trajectory baselines 3-4x and loses to persistence, so do not rank it by action L1
-
piper README: the held-out number that matters is the open-loop one (3.24 deg / 0.149), not the oracle-conditioned 0.093
-
README: the repo now also holds a real-robot policy
-
piper_pick_can_10k: interface, training budget, held-out numbers, and what is not established
-
piper_pick_can_10k: 20k phase-2 base + LoRA/flow/q4, 10k steps on the Piper pick-can data
14 days ago
-
Add libero_goal_full
-
Add libero_10_full
-
initial commit
