This is the memoryless control for the mu-VLA checkpoints: plain OpenVLA-OFT, with no memory tokens and no recurrent state, fine-tuned on the same five MIKASA-Robo-VLA environments, with the same episodic dataloader, for the same number of steps. It exists so that any gain reported by a mu-VLA checkpoint can be attributed to the memory module rather than to the data, the task mixture or the fine-tuning recipe.

The model is a per-step policy: every action is predicted from the current observation alone. On the memory-heavy MIKASA-Robo tasks, where the cue is shown early in the episode and is no longer visible when the action has to be taken, this is the behaviour a memory mechanism has to beat.

This checkpoint corresponds to experiment 2 in the mu-VLA experiment matrix, at training step 150000.

Sibling checkpoints, same five environments, with recurrent memory:

Base model: openvla/openvla-7b, fine-tuned with the OpenVLA-OFT recipe (L1 regression action head, proprioception projector, two input camera views, no FiLM, no diffusion). No memory module is present in the checkpoint.

The learning rate schedule is the difference to keep in mind when reading this model as a control: the memory checkpoints use a cosine schedule with 2000 warmup steps and a minimum ratio of 0.1, while this run uses the constant-then-decay schedule of the original OpenVLA-OFT recipe. Everything else, including the dataloader, the mixture and the number of steps, is identical.

The episodic dataloader feeds batch_size independent streams, each replaying whole episodes in order, and marks every step with is_first and is_last. It is used here as well, so that the control sees exactly the same frame ordering as the memory arms; without a memory state, that ordering only affects which frames land in a batch. Action normalisation statistics (q01/q99) are computed jointly over all five environments and stored under the mikasa_combined key in dataset_statistics.json.

Protocol: each environment is evaluated independently, 100 episodes, starting seed 4242424242, metric success_once.

Both inference modes are reported. Receding horizon queries the model at every environment step and executes only the first of the 8 predicted actions; open loop executes the whole predicted chunk before querying again. Receding horizon costs 8 times more forward passes and is the mode used for the numbers quoted in the memory checkpoints' cards.

Only the first five rows are training environments. Everything else is zero-shot transfer to environments this model never saw during fine-tuning.

There is no memory_module--150000_checkpoint.pt and no memory_meta.json here: that is the point of this checkpoint. Optimizer and scheduler state are not published for this run.

The checkpoint needs the mu-VLA code, which depends on a forked transformers, and on MIKASA-Robo for the environments. Point the evaluation script at the downloaded directory; with no memory_meta.json present it runs as plain OpenVLA-OFT. See CognitiveAISystems/muVLA for setup and the exact command line.

mu-VLA is a fork of OpenVLA-OFT by Moo Jin Kim, Chelsea Finn and Percy Liang (arXiv:2502.19645), which in turn builds on OpenVLA. The recurrent memory module, the episodic dataloaders and the MIKASA-Robo integration are the contribution of this work. Released under the MIT license, following upstream.