- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper proposes MemBodied, a fixed-size episodic memory for VLA models that combines a recurrent associative state, updated through a gated delta rule, with a fixed initial-scene anchor. Memory is updated with action–observation transitions and retrieved through a dedicated memory token. The method achieves substantial improvements on memory-dependent RMBench tasks and much lower inference latency than a compressed-history baseline (NativeMEM), with additional evaluations covering a π0.5 backbone, LIBERO, and real robots.
Critical strengths
The proposed fixed-size memory avoids growing observation histories, and the transition-conditioned update provides a sensible mechanism for retaining interaction information. The method outperforms stateless, frame-stacking, and recurrent baselines on RMBench, with lower latency than NativeMEM. The failure analysis provides useful insights, showing that memory improves decisions requiring information no longer available in the current observation.
Critical weaknesses
-
Evaluation appears to be insufficient. Only 5/9 of RMBench tasks are evaluated, and the benchmark’s own Mem-0 baseline is not included. Additionally, several comparisons seem questionable: the adapted µVLA baseline achieves 0% on every task, suggesting the adaptation failed rather than providing a meaningful comparison. The stateless π0 baseline (6.4%) falls below every published RMBench policy, suggesting a weak shared training recipe. The π0.5 and LIBERO improvements are measured against published rather than locally matched baselines. The real-robot evaluation is also limited: mean success rises from 3.3% to only 26.7%. Finally, the absence of multiple training seeds limits confidence in the results.
-
The mechanism largely transfers existing associative memory (the gated delta rule, Delta-Mem) to the VLA setting. The comparison with vanilla recurrent memory changes the write representation, readout interface, and anchor simultaneously, making it difficult to attribute the improvements to specific architectural choices.
Questions for authors
-
Why were only five RMBench tasks evaluated, and why was Mem-0 omitted? Can the authors show that the improvements persist against adequately optimized baselines under matched training conditions, particularly for π0.5 and LIBERO?
-
Can the authors isolate the contribution of transition-conditioned writing by comparing it directly against hidden-state writing, with the rest of the architecture unchanged?
-
For LIBERO, why does the memory update use the full predicted 50-step action chunk when only five actions are executed? How does this discrepancy affect training–inference consistency and performance?
2 comments
Response to Reviewer UnFW (Part 2)
Weakness 2:
The mechanism largely transfers existing associative memory (the gated delta rule, Delta-Mem) to the VLA setting. The comparison with vanilla recurrent memory changes the write representation, readout interface, and anchor simultaneously, making it difficult to attribute the improvements to specific architectural choices.
Response:
We adopt the gated delta update from Delta-Mem and adapt how memory is built for closed-loop control. Delta-Mem, designed for LLMs, writes from a text history available before prediction. Our VLA receives the current observation at each call, and the consequence of an action becomes available only after execution. We therefore build the state incrementally, writing each action chunk together with its resulting observation and training through the sequence so later action losses can shape earlier writes.
Vanilla Recurrent Memory and MemBodied, differ in the write, readout, and anchor. Table 1 includes two intermediate variants that let us examine those changes in sequence:
- Write representation. With attention steering and no anchor in both variants, MemBodied-AS writes the previous action and resulting observation in place of the hidden state. Success rises from 16.8% to 20.8%. This is the closest comparison for the write design, and the modest gain is suggestive.
- Readout. With transition-conditioned writes and no anchor, replacing attention steering with a dedicated memory token raises success from 20.8% to 37.6%. The token gives the action expert a representation of retrieved evidence that it can attend to directly.
- Anchor. Adding the initial-frame anchor to the memory-token variant raises success from 37.6% to 50.0%. The anchor keeps the starting scene available while the recurrent state tracks later interactions.
The largest gains in this sequence come from the readout and anchor. The write comparison measures the change under attention steering. A direct swap between hidden-state and transition-conditioned writes in the memory-token architecture would establish whether that gain carries over to the final readout.
Question 2:
Can the authors isolate the contribution of transition-conditioned writing by comparing it directly against hidden-state writing, with the rest of the architecture unchanged?
Response:
The closest comparison in Table 1 is Vanilla Recurrent Memory versus MemBodied-AS. Both use attention steering and omit the anchor. Success rises from 16.8% to 20.8% with transition-conditioned writes, which gives preliminary evidence in the attention-steering setting. Fig. 4b shows why we combine the two inputs. Across three tasks, writing the action together with its subsequent observation achieves 46.7% mean success, versus 42.7% for vision alone and 34.7% for action alone.
Question 3:
For LIBERO, why does the memory update use the full predicted 50-step action chunk when only five actions are executed? How does this discrepancy affect training–inference consistency and performance?
Response:
We replan every five environment steps in LIBERO, while the memory module was trained to update from observations spaced 50 steps apart. To preserve both cadences, we cycle through ten independent recurrent states. Each policy call uses one state.
For example, the first state is used at steps 0, 50, and 100, while the second is used at steps 5, 55, and 105. At step 50, the policy updates the first state with the new observation and uses the updated state to generate the next plan. Cycling through the ten states lets the policy replan every five steps while preserving the 50-step gap between successive observations used to update each state, matching the observation spacing used in training.
We hope our responses address your concerns. If they do, we would be grateful if you would consider updating your score. We are happy to answer any further questions.
Response to Reviewer UnFW (Part 1)
Dear Reviewer UnFW,
We appreciate your careful reading of our paper and your constructive comments. Please find our responses below.
Weakness 1/Question 1:
Response:
Task coverage and Mem-0. We chose five tasks to cover distinct recall decisions. Put Back Block, Rearrange Blocks, and Swap Blocks require recalling the original pad, block layout, and block–tray assignment, respectively. Battery Try and Block Ranking Try require tracking outcomes across attempts. In preliminary runs, two omitted tasks had near-zero success across the methods we tried. The other two were simpler or overlapped with the selected tasks. We omitted Mem-0 from this comparison because its separate planner maintains memory and feeds the actions to an execution module. Our local comparisons check how changing memory inside the action model affects the same VLA backbone.
Baselines. Our μ-VLA adaptation uses 64 recurrent tokens, sequences of eight samples at stride one, and gradient truncation every two recurrent steps, following the best configuration determined in the original paper. It achieved 0% success rate under the shared 10,000-step training budget. The original study trained OpenVLA-OFT on MIKASA-Robo for up to 150,000 steps. The local stateless π₀ shares the backbone, LoRA settings, and optimizer budget with the other local π₀ policies. Its 6.4% mean is low, but the paired rollouts help explain where the difference comes from. Both policies complete the early Put Back Block stages, and MemBodied corrects 24 wrong-pad choices with no reverse flips at the recall decision (Appendix E).
We also evaluated two additional baselines. Our π₀-MemoryVLA adaptation reaches 15.2% mean success and succeeds on all five tasks. To retain the initial scene needed by the three M(1) tasks, FrameStack (Initial) uses the first observation alongside those at t−30, t−15, and t. This gives the baseline access to both the initial scene and recent visual evidence. It reaches 34.0% mean success. Adding MemBodied raises this to 59.2%, with gains on four of five tasks, suggesting that Membodied can complement explicit frame history.
Matched comparisons. Matching MemBodied’s sequence training raises stateless π₀ from 6.4% to 9.6%. MemBodied reaches 50.0%, a 40.4 percentage-point gain, and leads on all five tasks. The largest gaps are on Rearrange Blocks (+78%) and Swap Blocks (+54%). Matched training may improve execution, but the policy still sees only the current observation at inference. It cannot reliably make the correct action from that observation alone for history-dependent manipulation.
On LIBERO-Long, MemBodied reaches 90.6% versus 88.0% for matched π₀. The advantage narrows from 5.4 points against published π₀ to 2.6 points against matched π₀. MemBodied also leads on Object (+0.6 points), while matched π₀ leads on Goal (+1.0) and Spatial (+2.4). The four-suite means remain nearly identical (95.10% versus 95.15%).
π₀.₅ comparison. The 3.87× figure compares MemBodied’s 48.0% with the published π₀.₅ result of 12.4%, showing that MemBodied can be applied to a second backbone. Table 1 and the matched π₀ controls reported above provide more direct evidence for the contribution of memory. On Put Back Block, for example, the correct return pad depends on the initial scene, which the current observation no longer reveals. Matching the training recipe could improve execution, but the stateless policy would still face that information constraint.
Training seeds. Repeating fine-tuning across the full set of baselines and ablations was computationally prohibitive within our budget. We used 50 shared evaluation seeds per RMBench task and also did a qualitative comparison at specific task stages in Appendix E. This controls for variation in evaluation and helps identify where MemBodied contributes.