Survey notes on World Action Models / VLA for robotics (index)

TL;DR

Luma-internal Google Doc index curated by Dong Guo and a co-author that organizes a set of group-by-group reading notes on representative World Action Model (WAM) and Vision-Language-Action (VLA) papers for robotics. The notes split the space into a few directions — explicitly video-as-policy and policy-first / data-flywheel approaches are called out — and prioritize the can and cannot of each direction: what it actually solves vs. only appears to, where each direction’s capability boundary sits, and how the directions connect and evolve. The artifact lives in Google Docs on purpose so the team can comment inline; the linked URL is the index, with individual direction notes branching out from there.

Key claims

• The survey organizes the WAM/VLA literature into directions rather than a flat paper list, and explicitly names video-as-policy and policy-first / data-flywheel as two of the directions covered [doc body, as described in the announcement message].
• Each per-direction note distills three things: the assumptions of the direction, its capabilities and limitations, and reusable ablation takeaways — the framing is orientation/positioning rather than implementation detail [doc body, as described in the announcement message].
• The notes prioritize identifying the capability boundary of each direction — what a direction actually solves vs. what it only appears to solve — and how the directions connect and evolve over time [doc body, as described in the announcement message].

Method

The artifact is a curated Google Doc index linking out to per-direction reading notes co-authored by Dong Guo and a teammate at Luma. The structure is group-by-group: each group is a “direction” in the WAM/VLA design space (with video-as-policy and policy-first / data-flywheel explicitly named), and each note within a group covers a representative paper with the assumptions / capabilities / limitations / ablation-takeaways pattern. The doc is editable by the whole team and is intended as a living document — inline comments are the suggested discussion mechanism. Access requires Luma SSO; the page body cannot be fetched anonymously, so this paper page records the announced structure rather than reproducing the notes themselves. A future /bud refresh after authenticated retrieval (or after the authors paste structured contents into a follow-up message) would let the page enumerate the specific directions and paper coverage.

Results

No quantitative results — this is a curated reading-list and synthesis document, not an experimental writeup. Its “result” is the taxonomy itself: a set of named directions plus an explicit “what each direction actually solves vs. only appears to solve” axis applied per paper.

Why it’s interesting

The most direct cross-link is Pretrained to Imagine, Fine-Tuned to Act: The Rise of World-Action Models — Moritz Reuss’ NVIDIA Technical Blog “State of VLA” / WAM taxonomy organized as a three-axis design space (paradigm × action integration × architecture). The Dong Guo survey appears to slice the same literature on a different axis (video-as-policy vs policy-first / data-flywheel directions, with explicit capability-boundary commentary per direction) and is a Luma-internal complement to the public Reuss taxonomy — the two together give the team both an architectural map (Reuss) and a “what does this actually solve” map (this survey). Also complements A Functional Taxonomy of World Models — World Labs’ three-way renderer / simulator / planner taxonomy operates at the purpose level, while this survey operates at the empirical “where does each direction break” level. For the active VLA Models concept page on the wiki — which currently tracks ~25 filed papers in the cluster and surfaces five competing recipes (action-pretraining-at-scale via π*0.6, clean-teleop-data via Spirit-v1.5, unified-VLM pointing via Embodied-R1.5, frozen-3D-trace-WM + small action expert via µ₀, unified causal flow-matching via LingBot-VA, plus the sensorimotor counter-position from Malik) — this survey is the most relevant internal synthesis filed so far and is the right starting point for anyone joining the cluster mid-stream. Note for future linking: when the doc is read in an authenticated context, individual paper notes likely correspond to entries already filed under VLA Models and World Foundation Models, and a refresh pass should backfill specific paper↔direction mappings.

See also

• VLA Models — the active concept page this survey is the most directly aligned with; ~25 papers tracking the five-recipe debate the survey organizes by capability boundary.
• World Foundation Models — the WFM cluster page; video-as-policy directions sit at the WFM↔policy boundary.
• Pretrained to Imagine, Fine-Tuned to Act: The Rise of World-Action Models — Moritz Reuss’ public WAM taxonomy (paradigm × action integration × architecture); the architectural counterpart to this survey’s direction-and-capability-boundary cut.
• A Functional Taxonomy of World Models — World Labs’ renderer / simulator / planner functional taxonomy of world models; the purpose-level counterpart.
• Robots Need More than VLA and World Models — Bou-Ammar position paper arguing VLA+WM is insufficient and four extra interfaces are needed; a direction-level critique the survey likely engages with.
• Direct Video-Action Models — Causal Video Models Are Data-Efficient Robot Policy Learners — DVA, the cleanest filed instance of the video-as-policy direction the survey explicitly names.
• Causal World Modeling for Robot Control (LingBot-VA) — LingBot-VA, a unified causal-flow-matching WAM that also belongs to the video-as-policy family.
• π*0.6: a VLA That Learns From Experience (RECAP) — π*0.6 / RECAP, the canonical policy-first / data-flywheel instance (closed-loop RL on real-robot data via advantage conditioning) the survey’s other named direction targets.
• Spirit-v1.5: Clean Data Is the Enemy of Great Robot Foundation Models — Spirit-v1.5, the strongest data-flywheel counter-bet at modest scale.