Real-Time Robot Execution with Masked Action Chunking

ListenThe Stale Prefix Problem in Real-Time VLAs9:54

Show notes

REMAC is a training-time recipe that makes flow-matching VLA policies robust under asynchronous inference — where the next action chunk is predicted while the current one still executes. The authors identify a second failure mode beyond the usual chunk-boundary discontinuity: intra-chunk inconsistency, where the executed prefix is stale relative to the current observation. They fix it with a LoRA fine-tune under a prefix-mask, self-conditioning curriculum, and residual-alignment loss, plus a prefix-preserved sampler — with zero added inference latency and gains that compose with BID and RTC. **Sources** - Real-Time Robot Execution with Masked Action Chunking

Subscribe to the channel: https://podcast.lumalabs.link/27a25f51ea81d41a8253f46c6133a35e/feed.xml

TL;DR

REMAC (Real-time Execution with Masked Action Chunking) is a training-time recipe for making flow-matching VLA policies robust under asynchronous inference — the deployment regime where the next action chunk is predicted while the current one is still executing. Previous methods (Bidirectional Decoding, RTC) attacked inter-chunk discontinuity at test time via inpainting or rejection sampling, but REMAC diagnoses a second, complementary failure mode — intra-chunk inconsistency, where the executed prefix of a chunk is stale relative to the current observation — and fixes it by fine-tuning the pretrained policy with LoRA (≤1.5% extra params) under a prefix-mask + self-conditioning curriculum + residual-alignment loss, plus a prefix-preserved sampler at inference. Reports higher success rates, faster completion, and better robustness across delays on 12 Kinetix tasks and three real-world Franka setups, and composes with existing test-time methods (BID, RTC) for further gains at no added inference latency.

Key claims

Two failure modes, not one. Asynchronous inference under action chunking causes both inter-chunk discontinuity (mode-switches at chunk boundaries) and intra-chunk inconsistency (the delayed prefix of chunk k+1 was conditioned on the wrong observation o_t instead of o_{t+δ}); prior work (TE, BID, RTC) only addresses the former [§3].
Prefix masking during flow-matching training restricts the loss to the to-be-executed suffix of each chunk, with the mask cut point sampled uniformly across all valid delays δ so a single fine-tuned policy handles the full delay spectrum without retraining per delay setting [§4.1, Eq. 2–3].
Self-conditioned curriculum linearly anneals training inputs from ground-truth action chunks to samples drawn from the pretrained policy (with stop-gradient), which the paper argues aligns training-time inputs with test-time conditions and mitigates exposure bias while keeping early-training stable [§4.1, Eq. 4].
Residual-alignment loss explicitly supervises the LoRA correction to match the residual between the pretrained flow estimate and the ground-truth target, in addition to the direct ground-truth loss — the ablation table shows both terms are load-bearing [§4.1, Eq. 5–6].
Prefix-preserved sampling initializes the flow integration from the previously executed actions (rather than Gaussian noise) and freezes those positions across all integration steps, aligning inference with the training-time mask [§4.2, Eq. 7].
No added inference latency. LoRA modules can be merged into the backbone at deployment, so REMAC costs zero extra forward-pass time relative to the pretrained π₀-style policy [§4.3].
Kinetix results (12 tasks, δ ∈ 0…4). REMAC consistently outperforms Naive Async, BID, and RTC across all delay settings, with the largest margins at higher delays; also improves at δ=0, suggesting the training recipe strengthens temporal coherence independently of delay [§5.1, Fig. 2].
Composable with test-time methods. Combining REMAC with BID or RTC yields further additive gains at higher delays [§5.1, Table 2].
Ablation isolates the training recipe. LoRA alone with unchanged loss delivers no gain, confirming that the improvement is not merely from extra parameters but from the prefix-mask + curriculum + residual terms [§5.1, Table 3].
Extends beyond flow matching. The same recipe applied to the Transformer-based ACT framework beats both naive async and LoRA-only baselines [Appendix E.5].

Method

Given a pretrained flow-matching action-chunking policy π̄ (e.g. π₀-style), REMAC learns a small LoRA correction (rank 4 on all linear layers except time-embedding and AdaLN) under three combined losses on the unmasked suffix of each chunk: (1) a masked flow-matching loss against the ground-truth target restricted to timesteps beyond the delay cut δ, (2) a self-conditioned curriculum where the interpolation endpoint anneals from the ground-truth action chunk to a stop-gradient rollout from the pretrained policy π̄, and (3) a residual-matching term that aligns the LoRA correction with the gap between the pretrained flow and the ground-truth target. δ is sampled uniformly from a training-time schedule that shrinks a [0, δ_max] interval over training. At inference, REMAC’s prefix-preserved sampler fills the first δ entries of the initial latent with the already-executed actions, zero-pads the rest, and freezes those δ positions across all integration steps of the learned velocity field. Merging LoRA into the backbone keeps deployment latency identical to the base policy.

Results

Kinetix (12 tasks, RPO-generated expert data, IL training, prediction horizon 8): REMAC beats Naive Async, BID, and RTC on both success rate and completion time across δ ∈ 0…4; gap widens with δ [Fig. 2].
Composition: +REMAC on top of BID and RTC gives further modest gains, larger at higher delays [Table 2].
Ablations [Table 3]: LoRA-only ≈ baseline; adding prefix masking → gain; adding self-conditioning curriculum → further gain; adding residual alignment → full method, highest overall SR.
Curriculum schedule [Fig. 3]: piecewise-linear annealing from α=0 (ground-truth) to α=1 (self-conditioned) beats both constant α=0 and constant α=1 in success rate and training stability.
Real-world Franka Research 3 (DROID setup, three settings): higher completion rates, faster task times, smoother trajectories under varying network + inference delays; grippers with narrowed range used to stress-test fine-grained control [§5.2].
Extends to ACT (Transformer-based): consistently beats Naive Async and LoRA-only baselines [Appendix E.5].
Hyper-parameter sensitivity [Fig. 10a]: performance mostly flat across δ_min / δ_max; larger δ_max slightly worse; authors use δ_min=1, δ_max=3.

Why it’s interesting

This is the sharpest filed treatment of the inference-delay half of the VLA-deployment problem, and it slots cleanly onto the recipe-lever board on VLA Models as a training-time recipe for async-inference robustness that is orthogonal to (and composable with) the test-time inpainting recipe of RTC and the test-time rejection-sampling recipe of BID. The paper’s diagnostic split — intra-chunk inconsistency vs inter-chunk discontinuity — is the useful conceptual contribution: prior work implicitly conflated the two under “chunk boundary smoothing”, but the executed-prefix’s staleness relative to the current observation is a distinct problem that test-time inpainting cannot fix without extra compute. REMAC’s prefix-preserved sampler also has a suggestive resemblance to the async pipeline in Causal World Modeling for Robot Control (LingBot-VA) (LingBot-VA’s FDM-grounded re-grounding against the freshest observation) and to A2A: Action-to-Action Flow Matching‘s A2A (which shortens the flow trajectory itself by anchoring one endpoint on the previous proprioceptive action) — three different ways of exploiting the previous chunk / previous observation as a structured prior for the next inference step. Compared to A2A’s aggressive 1-step Euler goal, REMAC keeps the standard multi-step flow-matching sampler and pushes the fix into the training objective, which is why it slots on top of BID/RTC without additional inference cost.