rldx-1-test

RLDX-1 fine-tuned checkpoint for Dongkkka/Task_99999_pick_place_snack_cut_processed.

This repository contains inference artifacts only:

• model config and sharded safetensors weights
• processor/ files for modality config, embodiment mapping, and normalization statistics

It intentionally excludes optimizer state, scheduler state, trainer state, RNG state, training logs, and resume checkpoints.

Training Summary

• Base model: RLWRLD/RLDX-1-PT
• Dataset: Dongkkka/Task_99999_pick_place_snack_cut_processed
• Camera input: observation.images.rgb.cam_left_head only
• Task text: pick up yellow snack and place it into gray box.
• Steps: 2000
• Global batch size: 208
• RTC training max delay: 4
• Final train loss: 0.006251013543456793

Expected RLDX Modality

• Video keys: cam_left_head
• Video delta indices: [-6, -4, -2, 0]
• State delta indices: [0]
• Action horizon: 16

Minimal Usage

from rldx.data.embodiment_tags import EmbodimentTag
from rldx.policy.rldx_policy import RLDXPolicy

policy = RLDXPolicy(
    embodiment_tag=EmbodimentTag.GENERAL_EMBODIMENT,
    model_path="Dongkkka/rldx-1-test",
    device="cuda:0",
    strict=True,
)

Run from a compatible RLDX-1 environment with the RLDX model classes registered.