research-vla/RIFT-RoboCOIN

RIFT on RoboCOIN (three tasks)

RIFT trained for ten epochs on three RoboCOIN datasets, in two variants that differ only in how the two training paths are batched during training.

Variant	model.fuse_training_paths	Seconds / optimizer step	Ten epochs
fused_z1_10ep	true	3.550	9.21 h
separate_z1_10ep	false	3.800	9.86 h

Both ran on two nodes of four H100 94 GB cards under DeepSpeed ZeRO-1, bf16, batch 16 per rank with two accumulation steps, for a global batch of 256 and 9,340 optimizer steps.

The fused variant batches the prepared tokens of both training paths into one MoT call. It changes no module and no parameter, so the two checkpoints share the same state_dict keys and shapes and load into the same inference code without any branch. Both runs kept model.train_probe_head at its default true; that flag does change the parameter set, and a checkpoint saved with it disabled loads through strict=False without warning, leaving the probe randomly initialized.

Data

342 episodes and 238,864 windows across RoboCOIN/Galaxea_R1_Lite_classify_object_four (191 episodes), RoboCOIN/Galaxea_R1_Lite_mix_red_yellow_large_test_tube (50) and RoboCOIN/R1_Lite_stack_baskets (101), one instruction each. Training read the Galaxea-compatible view: left arm 6, left gripper, right arm 6, right gripper, grippers on the 0-100 millimetre stroke, cameras aliased to head_rgb, left_wrist_rgb and right_wrist_rgb.

Files per variant

• <variant>/rift_stepXXXXXX.pt — weights only, saved with optimizer=None.
• <variant>/config.yaml — the resolved training configuration.
• <variant>/dataset_stats.json — normalization statistics. Inference needs these and matching field and unit preprocessing.