Most progress in vision-language models comes from scaling perceptual data, but robotics is bottlenecked by scarce, costly robot trajectories. The core insight here is that with a fixed robot-data budget, improving the quality and cross-embodiment alignment of representations matters more than naive data scaling: a VLA backbone that preserves VLM priors while explicitly shaping action semantics can convert limited trajectories into broadly transferable control knowledge.
Key Findings
- Representation-centric continued pre-training (VLAct) improves downstream VLA transfer under fixed fine-tuning protocols, outperforming several industrial VLA systems on standard benchmarks.
- Main mechanisms: VLM-prior preservation (keeps broad visual-language capabilities), multi-head continuous-action co-supervision (encourages shared action semantics), and a partially unified cross-embodiment action layout (enables transfer while allowing task-specific heads at fine-tuning). Together these reduce overfitting to narrow robot data and improve generalization to unseen embodiments.
- Empirical benefits include higher success rates on LIBERO-Plus and RoboTwin 2.0, competitive ranking on RoboDojo, and strong few-shot transfer to unseen humanoid embodiments (notably outperforming a full-data baseline using only 20% of downstream trajectories).
- Achieves these gains with fully open-source data and a modest 16-GPU training setup, illustrating practical cost-efficiency compared to pure data-scaling approaches.
How it Works (brief)
VLAct begins from a vision-language backbone and continues pre-training on broad, heterogeneous multi-embodiment robot data with objectives that (1) preserve the VLM prior to retain language-grounded perception, (2) supervise continuous action embeddings with multiple heads to capture both shared and embodiment-specific signals, and (3) adopt a partly unified action layout so that cross-embodiment semantics are learnable while fine-tuning can attach task-specific action heads.
Who it's for and trade-offs
Great fit if you need stronger transfer from limited robot trajectories, care about cross-embodiment generalization, and want competitive VLA performance without massive specialized robot-data scaling. Look elsewhere if you can collect large-scale, high-quality task-specific robot action datasets or require methods optimized solely for a single fixed embodiment; VLAct focuses on representation and transfer rather than maximizing single-robot asymptotic performance.