UC Riverside

Vision–Language–Action (VLA) models have become a dominant paradigm for generalist robot manipulation, yet most are effectively memoryless: they map the current observation and a language instruction to an action without an explicit mechanism for retaining and reusing information across long-horizon, partially observable episodes. This thesis makes two contributions toward memory-augmented VLA policies. First, we unify three heterogeneous VLA backbones—pi0.5, Cosmos, and DreamZero— under a single BasePolicy interface, so the same evaluation and rollout infrastructure applies uniformly to all three. Second, we introduce DynaMem, an input-level, single-point dynamic memory injection mechanism that augments any backbone behind this interface without modifying its internal architecture or requiring backbone-specific engineering. We evaluate the resulting system with closed-loop rollouts on the RoboMemArena (RMA) benchmark. Preliminary experiments indicate that input-level memory injection improves task completion on memory-dependent manipulation tasks while maintaining compatibility with standard evaluation pipelines.Our experiments yield a careful negative result: when DynaMem is fine-tuned on top of a converged behavior-cloned π0.5 policy, task performance is statistically indistinguishable from a no-memory control, and zeroing the memory at inference time does not degrade performance—the policy learns to ignore the injected tokens. We trace this to a structural property of behavior cloning from plateaued checkpoints: the training loss can be minimized without consulting memory, so the zero-initialized memory pathway receives no gradient pressure. The injection overhead itself is negligible (under 10 ms per memory write and 12 µs per control step). We distill these findings into design guidelines for future memory-augmented VLA training.