mu-VLA extends OpenVLA-OFT with a recurrent memory module so that the policy can act in partially observable Markov decision processes (POMDPs), where the information needed for the current action appeared earlier in the episode and is no longer visible. Vanilla OpenVLA-OFT is a per-step policy and cannot solve such tasks.

Memory is implemented as a set of learnable memory tokens that are injected into the multimodal prefix of the transformer sequence and carried across environment steps. Gradients flow through a truncated backpropagation-through-time (TBPTT) window; the memory state is reset at episode boundaries.

This checkpoint corresponds to experiment 12 in the mu-VLA experiment matrix: multitask training on 5 MIKASA-Robo-VLA environments, 64 memory tokens, TBPTT window K=2. Of the mu-VLA MIKASA-Robo checkpoints released so far, this is the strongest one on the memory-intensive environments.

Sibling checkpoints:

The same values are shipped in memory_meta.json so that the evaluation scripts in the code repository pick them up automatically.

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).

The episodic dataloader feeds batch_size independent streams, each replaying whole episodes in order, and marks every step with is_first and is_last so that the memory state can be reset on episode boundaries. 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. Uncertainty is the standard error over 100 Bernoulli trials.

Inference uses receding horizon: instead of executing the whole open-loop chunk of 8 predicted actions, the model is queried every step and only the first action is executed. This matters for POMDP tasks where the visual cue is visible only briefly. The cost is 8 times more forward passes. On this checkpoint the effect is large; the open-loop column below is the same weights evaluated without receding horizon.

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

Baseline column: OpenVLA-OFT without memory, trained on the same five environments with the same episodic dataloader (experiment 2 in the matrix). It is not a perfectly matched control, since it used the constant-then-decay learning rate schedule rather than cosine.

The Rotate* and InterceptGrab* families stay near zero. Those require precise contact-rich control rather than recall, so memory does not help there; they are reported for completeness.

Loading this checkpoint requires more than the base transformer weights. The extra heads are stored as separate PyTorch state dicts, matching the layout the training and evaluation scripts expect.

The checkpoint is not loadable with upstream transformers. It needs the mu-VLA code, which depends on a forked transformers that supports the custom attention mask, and on MIKASA-Robo for the environments.

Then point the evaluation script at the downloaded directory; memory hyperparameters are read from memory_meta.json. See the code repository, CognitiveAISystems/mu-vla, 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.