LingBot-VA AgileX pour β€” RoboMeter RL checkpoints

The RL20, RL50, RL100 and RL200 directories contain the PEFT adapters from updates 20, 50, 100 and 200 of rbm-rl-pour-384-1110-r1. Each contains its original adapter_config.json and adapter_model.safetensors (480 FP32 tensors, rank 32, alpha 64).

Use the base transformer at the evaluation revision. Attach one adapter to the transformer and enable it for video prediction. The evaluated action branch uses the same frozen base transformer with the adapter disabled. Preserve the FP32 adapter parameters when using a BF16 base.

To download a single checkpoint from this repository with an authorized Hugging Face account:

from huggingface_hub import snapshot_download
from pathlib import Path

checkpoint = "RL50"  # RL20, RL50, RL100 or RL200
root = snapshot_download("MarshCurrant/lingbot-va-agilex-pour-robometer-rl", allow_patterns=[f"{checkpoint}/*"])
adapter_dir = Path(root) / checkpoint

The four adapters were evaluated on the same 20 held-out initial observations from the AgileX pouring task. Visible completion at the original episode duration was 19/20 for RL20, 16/20 for RL50, 0/20 for RL100 and 0/20 for RL200. The latter two produced sustained visual collapse in all 20 imagined rollouts. These are video rollout measurements without hardware trials.

RL50 multi-task IDM fine-tuning

RL50_FT_IDM contains the validation-selected IDM step 950: a native transformer, exact FP32 action parameters, unchanged RL50 video adapter and matching four-task normalization. It is an experimental variant, not a replacement for the original adapters. Paired test80 action reconstruction did not improve for mixed-pour/RL20 or other task-specific checkpoints; see the release's evaluation tables and limitations. Existing RL20/RL50/RL100/RL200 files are unchanged.

Playdata + SFT run, 2026-11-16

playdata_sft_1atyoo3a/RL20 contains the current-policy update-20 adapter from run 1atyoo3a. It uses the new playdata + SFT base and its task+play normalization; see the folder README and provenance. The root-level RL adapters and their evaluation descriptions above belong to the earlier run.

Downloads last month
-
Video Preview
loading

Model tree for MarshCurrant/lingbot-va-agilex-pour-robometer-rl

Adapter
(1)
this model