k3rth1v5n/forcebench-sim-models

ForceBench sim models (final protocol)

Final checkpoints for the five Genesis simulation tasks, trained with measured-TCP ("tip") labels: diffusion policies for 300 epochs, VLAs for 15 epochs. Push box uses the 75-episode recording.

<task>/afp_300ep/        latest.pt, normal_force_law.pt, config.yaml
<task>/acp_300ep/        latest.ckpt, sparse_normalizer.pkl, config.yaml
<task>/maniflow_300ep/   model_epoch_0300.pt, config.yaml
<task>/rdp_300ep/ldp/    latest.ckpt, normalizer.pkl, config.yaml   (latent diffusion policy, 300 epochs)
<task>/rdp_300ep/at/     latest.ckpt, config.yaml                   (asymmetric tokenizer, 601 epochs)
<task>/forcevla_15ep/    params/, assets/, _CHECKPOINT_METADATA     (batch 32, ceil(15 * frames / 32) steps)
all5/molmoact2_15ep/     step11501-merged/ (serve this), step11501-lora-llm/, step11501-lora-vision/, config.yaml

<task>: pivot_block, plug_insertion_round, plug_insertion_usb, bottle_wiping, push_box. MolmoAct2 is one LoRA fine-tune of MolmoAct2-DROID on all five tasks together (tag forcebench_<task> per task, global batch 64, 11,501 steps).

Task	Episodes	Frames
pivot_block	50	12,328
plug_insertion_round	50	8,002
plug_insertion_usb	50	9,080
bottle_wiping	65	12,333
push_box	75	7,325