SLA: Scaling Spatial-Language-Action Model via Continual Pre-training

TL;DR

SLA is a Spatial-Language-Action Model initialized from a pretrained π0.5 policy and continually pre-trained to inject 3D spatial understanding into a VLA foundation model, by aligning intermediate visual embeddings against spatial representations from a frozen 3D perception teacher. The authors argue that today’s VLAs inherit visual representations from 2D vision-language pre-training and only acquire spatial grounding during task-specific adaptation — a scale mismatch that SLA closes by treating spatial alignment as a transferable foundation-model capability learned on a unified heterogeneous robot corpus. The headline deployment claim (from the Slack pointer) is top-of-leaderboard performance on RoboDojo’s real-world track; the paper’s own technical signatures are feature-dependent gating that pools visual embeddings across multiple VLA layers for spatial supervision, a depth-plus-calibration teacher that works on partially-annotated geometric data, and probabilistic state dropout to prevent trajectory memorization — all auxiliary components dropped at inference so the original π0.5 architecture is preserved.

Key claims

Vision-language-action foundations largely inherit 2D vision-language representations and introduce spatial supervision only at task-specific adaptation time, which misallocates where spatial grounding is learned [Abstract].
Spatial understanding should be developed as a transferable foundation-model capability via continual pre-training from broader robot experience, not re-derived per task [Abstract].
SLA is initialized from a pretrained π0.5 policy and jointly optimizes action prediction and alignment between intermediate VLA visual embeddings and spatial representations from a frozen 3D perception model [Abstract].
Feature-dependent gates adaptively aggregate visual embeddings from multiple VLA layers for spatial supervision, to accommodate embodiment / task / scene diversity in the pretraining corpus [Abstract].
The 3D teacher exploits partially-available depth and camera-calibration channels alongside RGB — the training setup does not require every sample to carry full geometric annotation [Abstract].
Probabilistic state dropout encourages reliance on visual cues and mitigates overfitting to demonstration trajectories [Abstract].
All auxiliary components (spatial alignment loss, gates, teacher) are used only during training; pretrained VLA weights are reused directly at inference and the original architecture is preserved [Abstract].

Method

SLA is a continual pre-training recipe on top of the pretrained π0.5 VLA, not a from-scratch architecture. During pre-training, each sample produces (a) the standard action-prediction objective π0.5 already has, and (b) an auxiliary spatial-alignment objective that pulls intermediate visual embeddings inside the VLA toward spatial representations emitted by a frozen 3D perception teacher. To make this alignment work across embodiments, tasks, and scenes, SLA adds feature-dependent gates that adaptively mix visual embeddings from multiple VLA layers before alignment — a token-wise router per sample rather than a single fixed layer. The teacher itself is designed around the dataset reality: some samples have RGB + depth, some have RGB + camera calibration, few have both; the teacher consumes whatever geometric annotation is present alongside RGB, so a unified heterogeneous robot corpus can be used without filtering down to the fully-annotated subset. On the input side, probabilistic state dropout randomly suppresses robot-state inputs during training, forcing the action head to rely more on visual cues and less on trajectory memorization. At inference, the 3D teacher, the gates, and the spatial-alignment loss are all removed — the deployed model is the π0.5 architecture with pretrained weights continued on this objective, so existing π0.5 serving stacks apply unchanged.

Results

The arxiv abstract retrieved at filing time does not report concrete benchmark numbers (RoboDojo-sim / RoboDojo-real / LIBERO / LIBERO-Plus / RoboTwin / task-level success rates); the only headline performance claim available comes from the Slack pointer, which states that the authors claim to “reach the top of RoboDojo real-world eval.” This slot is flagged for /bud refresh once the full PDF with tables is parsed — numeric baselines, per-task breakdown, π0.5 head-to-head, and ablations on (i) the feature-dependent gates, (ii) partial-annotation teacher usage, and (iii) probabilistic state dropout are the loads-bearing numbers that should land here.

Why it’s interesting

SLA stakes out a distinct where-to-learn-spatial-grounding position on the VLA Models recipe-lever board. Where Embodied-R1.5: Evolving Physical Intelligence via Embodied Foundation Models argues a strong pointing/grounding-pretrained VLM backbone substitutes for action pretraining, SLA keeps π0.5’s action-pretrained backbone and retrofits spatial grounding via continual pre-training against a frozen 3D teacher — the closest filed sibling in that cell is Anchor-Align: Generalizable VLA Finetuning via Representation Anchoring and Language-Action Alignment Anchor-Align, which also uses a frozen teacher to anchor representations during a continued-training stage, but Anchor-Align distills the VLA’s own frozen VLM against itself to prevent forgetting, while SLA distills an external 3D perception model to inject new geometric structure. Both share the “representation-shaping auxiliary objective, discarded at inference” pattern that Beyond Data Scaling: Representation-Centric Continued Pre-training for Vision-Language-Action Models VLAct formalizes most aggressively (freeze lower half of the LLM, discard action heads after pretraining). SLA also sits adjacent to Grounded Action Model: 3D Grounding as a Foundation for Robotics GAM, which argues 3D grounding should be the primary pretrained substrate rather than a continual-pre-training retrofit on top of a VLM — SLA is the “retrofit” counter-position to GAM’s “ground-up” position. If the RoboDojo real-world claim holds, that would be the first filed instance of a continually-pre-trained π0.5 variant topping a third-party real-world leaderboard where RoboDojo: A Unified Sim-and-Real Benchmark for Comprehensive Evaluation of Generalist Robot Manipulation Policies places 30 integrated policies on a shared protocol — directly comparable to Embodied Turing Machines: Stateful Code for Robot Recursive Self-Improvement COAP’s 70.24% RoboDojo-sim result (no model at test time) and to Democratizing Robot-Use Agents: A Unified Harness with Open Models uni-robot-agent’s 31.39 RoboDojo score with an open LLM controller.