uber0802/CRONOS-spatialvla-T2560-normal-EER

CRONOS SpatialVLA — T2560 normal (EER)

LoRA adapters (rank 32) on top of IPEC-COMMUNITY/spatialvla-4b-224-sft-bridge, trained with PPO in CRONOS on ManiSkill PickPlaceNxM-v1 (2x2, four_group_sequential_2x2).

These are the final checkpoints (episode_0004) of two seeds.

Layout

seed0/episode_0004/
seed1/episode_0004/
    adapter_model.safetensors   # LoRA adapter weights
    adapter_config.json         # PEFT config (base class: SpatialVLAForActionPredictionWithValueHead)
    dataset_statistics.json     # action normalization stats (unnorm key: bridge_orig/1.0.0)
    training_state.pt           # value head + VLA/value-head optimizer states + step counters (for resuming)
    scheduler_state.json        # task scheduler state
    run_config.yaml             # full training config
    README.md                   # PEFT auto-generated card

Training setup

Env	PickPlaceNxM-v1, 2x2, 64 envs, obj_set: rand
Task order	sequential, 4 groups, 16 tasks
Episode length	2560 steps (segment / task length 80)
Reset	per_episode, reset_robot: true (EER)
Episodes	4 (655,360 total env steps)
Algorithm	PPO, 1 epoch, grad accum 20, gamma 0.99, lambda 0.95
LR	VLA 1e-4, value head 3e-3, grad norm 10
LoRA	r=32, alpha=16
Seeds	0, 1
CRONOS	V0.3, git rev e45cdb86f5cb717413258a1713bab07820fc9ea0