CollabVLA: Self-Reflective Vision-Language-Action Model Dreaming Together with Human

The gist The authors present CollabVLA, a self-reflective vision–language–action framework that transforms a standard visuomotor policy into a collaborative assistant.

How it works

CollabVLA tackles key limitations of prior VLAs, including domain overfitting, non-interpretable reasoning, and the high latency of auxiliary generative models by integrating VLMbased reflective reasoning with diffusion-based action generation under a mixture-of-experts design<ref:2509.14889#pg2>. It follows a two-stage training recipe: (1) Action Grounding, where a VLM-driven action policy is trained on latent action representations conditioned on multimodal goals to master acting, and (2) Reflection Tuning, which unifies scene understanding and action generation while strengthening robust internal reflective reasoning<ref:2509.14889#pg2>.

The problem formulation defines CollabVLA as a goal-conditioned vision–language–action policy that outputs not only an action sequence but also explicit reflective reasoning and optional human queries<ref:2509.14889#pg2>. Given a current state o t (image observations), past states o t−1, proprioception p t, and a multimodal goal g, the policy πθ predicts: πθ(o t, ot−1, pt, g) → a¯t, rt, qt<ref:2509.14889#pg2>. The output components are defined as follows: a¯t = (a t i,..., a t i+k−1) denotes a short action chunk, r t is a reflective reasoning trace, and q t is a binary query<ref:2509.14889#pg2>.

Model Architecture

CollabVLA couples a VLM backbone with MoE adaptation and a diffusion-based action expert for conditional, low-latency trajectory generation<ref:2509.14889#pg2>. The VLM Backbone builds on InternVL2.5, consuming a single sequence that concatenates robot observations tagged as “[NOW]” and “[PAST],” cached embeddings for o t−1, a multimodal goal g plus optional human guidance wrapped by “…,” proprioception p t projected by a small MLP, and K learnable [ACT] queries<ref:2509.14889#pg2>. The model outputs (a) a reflection string, (b) a binary ask indicator from a classifier over the pooled reflection representation, and (c) K latent action embeddings from the final hidden states of the [ACT] queries<ref:2509.14889#pg2>.

Inside the backbone, a MixtureofExperts design enables alternation between reflection and forward planning through an adaptive gating scheme similar to InstructVLA<ref:2509.14889#pg2>. LoRA experts are inserted into the linear projections of MHA and FFN, specifically a Control Expert and a Reflection Expert. The output for hidden state x is given by: h = W0x + (BctrlActrlx)αctrlλctrl + (BrefArefx)αrefλref<ref:2509.14889#pg2>. This adaptive gating favors the Control Expert during routine control and shifts to the Reflection Expert under uncertainty, allowing the model to balance reasoning and acting<ref:2509.14889#pg2>.

The Diffusion-Based Action Model employs a Diffusion Transformer (DiT) as the action generator<ref:2509.14889#pg2>. The VLM backbone provides two signals to this DiT: (i) latent action tokens from learnable [ACT] queries, which are injected as key–value memories in cross-attention for fine-grained trajectory control, and (ii) a reflection embedding, the final hidden states of reflection tokens, broadcast via FiLM to modulate hidden activations with global semantic guidance<ref:2509.14889#pg2>.

Training Pipeline

The training employs a two-stage recipe: Action Grounding and Reflection Tuning<ref:2509.14889#pg2>. In Stage 1, Action Grounding aligns multimodal perception with motor actions and lightweight planning language, activating only the Control Expert. The overall loss is LStage1 = λlang · Llang + λact · Ldiff<ref:2509.14889#pg2>. Here, Llang updates the Control-Expert LoRA θctrl, while Ldiff updates the DiT parameters ϕ and θctrl<ref:2509.14889#pg2>.

In Stage 2, Reflection Tuning freezes the diffusion action model and VLM backbone while jointly training the two VLM-side LoRA experts and auxiliary heads. The objective is to unify action generation with reflective reasoning when encountering failures or ambiguities<ref:2509.14889#pg2>. The overall training loss is LStage2 = λref · Lref + λask · Lask<ref:2509.14889#pg2>. Here, Lref updates (θctrl, θrefl, Wgate), and Lask updates only the ask head ψ<ref:2509.14889#pg2>.

Experimental Evaluation

CollabVLA was evaluated across four questions regarding multimodal understanding preservation, ambiguity detection and reflection quality, interpretation of human-in-the-loop feedback for long-horizon tasks, and efficiency with minimal overhead<ref:2509.14889#pg2>. Empirical results show that CollabVLA matches/exceeds strong VLAs and remains competitive with larger MLLMs across comprehensive understanding and VQA benchmarks<ref:2509.14889#pg2>.

In simulation experiments, CollabVLA achieves the best success rate in SimplerCollab across all 8 subtasks (Table II)<ref:2509.14889#pg2>. Against the strongest VLA baseline (InstructVLA), CollabVLA improves the mean by +4.8/+1.6/−2/+0.3/+26.6 on the above metrics<ref:2509.14889#pg2>.

In real-world experiments, CollabVLA achieves the best average SR/Score (2.5/62.2), leading in 4/5 tasks and showing superior ability to interpret multimodal human guidance compared to OpenVLA–Collab<ref:2509.14889#pg2>. The results underscore that CollabVLA’s reflect–ask/act loop with FiLM–conditioned control turns brief human hints into more reliable progress under clutter, ambiguity, and multi–step objectives<ref:2509.14889#pg2>.

Conclusion

CollabVLA introduces a self-reflective VLA that reasons explicitly, reflects on uncertainty, and integrates lightweight human feedback in real time<ref:2509.14889#pg2>. The two-stage recipe yields consistent gains in success and interpretability across tasks<ref:2509.14889#pg2>. The work demonstrates that CollabVLA is a practical, collaborative assistant that unifies perception, action, and collaboration<ref:2509.14889#pg2>. Looking ahead, future avenues could include integrating tactile and force sensing or evolving collaboration from queries to proactive task allocation<ref:2509.14889#pg2>.

--- Page 5 ---

The authors identify and systematize the trade-offs among (a) direct autoregressive VLAs, (b) latent-action formulations, and (c) explicit world-model approaches, highlighting the missed opportunity for lightweight human-in-the-loop guidance at execution time<ref:2509.14889#pg2>. The paper introduces CollabVLA, a collaborative VLA framework with MoE adaptation that transforms a standard visuomotor policy into a proactive assistant capable of reasoning, acting, and interacting with humans<ref:2509.14889#pg2>. We show that CollabVLA improves success rates and maintains low latency while effectively extending its self-reflection to solicit just-in-time human guidance<ref:2509.14889#pg2>. This work takes a pioneering step toward shifting VLAs from opaque controllers to genuinely assistive agents capable of reasoning, acting, and collaborating with humans<ref:2509.14889#pg2>. The authors identify and systematize the trade-offs among (a) direct autoregressive VLAs, (b) latent-action formulations, and (c) explicit world-model approaches, highlighting the missed opportunity for lightweight human-in-the-loop guidance at execution time<ref:2509.14889#pg2>. We introduce CollabVLA, a collaborative VLA framework with MoE adaptation that transforms a standard visuomotor policy into a proactive assistant capable of reasoning, acting, and interacting with humans<ref:2509.14889#pg2>. We show that CollabVLA improves success rates and maintains low latency, and effectively extends its self-reflection to solicit just-in-time human guidance<ref:2509.14889#pg2>. This work takes a pioneering step toward shifting VLAs from opaque controllers to genuinely assistive agents capable of reasoning, acting, and collaborating with humans<ref:2509.14889#pg2>. The authors identify and systematize the trade-offs among (a) direct autoregressive VLAs, (b) latent-action formulations, and (c) explicit world-model approaches, highlighting the missed opportunity for lightweight human-in-the-loop guidance at execution time<ref:2509.14889#pg2>. We introduce CollabVLA, a collaborative VLA framework with MoE adaptation that transforms a standard visuomotor policy into a proactive assistant capable of reasoning, acting, and interacting with humans<ref:2509.14889#pg2>. We show that CollabVLA improves success rates and maintains low latency, and effectively extends its self-reflection to solicit just-in-time human guidance<ref:2509.14889#pg2>. This work takes a pioneering step toward shifting VLAs from opaque controllers to genuinely assistive agents capable of reasoning, acting, and collaborating with humans<ref:2509.14889#pg2>. The authors identify and systematize the trade-offs among (a) direct autoregressive VLAs, (b) latent-action formulations, and (c) explicit world-model approaches, highlighting the missed opportunity for lightweight human-in-the-loop guidance at execution time<ref:2509.14889#pg2>. We introduce CollabVLA, a collaborative VLA framework with MoE adaptation that transforms a standard visuomotor policy into a proactive assistant capable of reasoning, acting, and interacting with humans<ref:2509.14889#pg2>.

1. Bold header: Incorporate self-reflective reasoning into action generation to create a collaborative assistant. This allows agents to move from opaque controllers to genuinely assistive agents capable of reasoning, acting, and collaborating with humans, proactively soliciting human guidance when facing uncertainty or repeated failure.

2. Bold header: Implement an adaptive gating scheme using Mixture-of-Experts (MoE) for the VLM backbone that alternates between a Control Expert and a Reflection Expert. This design ensures the model can balance reasoning and acting, as described by the formula where The adaptive gating favors the Control Expert during routine control and shifts to the Reflection Expert under uncertainty.

3. Bold header: Develop a two-stage training recipe consisting of Action Grounding followed by Reflection Tuning. The first stage trains a VLM-driven action policy on latent action representations, while the second stage unifies scene understanding and action generation, thereby maintaining policy performance while strengthening robust internal reflective reasoning.

4. Bold header: Integrate a binary query mechanism, where the policy predicts an ask indicator q t ∈ [0, 1] specifies whether to solicit human input, allowing the agent to be a reflective collaborator that reports reasoning outcomes and proactively incorporates concise textual hints or lightweight visual cues from human input.

5. Bold header: Utilize a diffusion-based action generator conditioned on both latent action tokens and a reflection embedding modulated via FiLM. This ensures that the generated actions are dynamically consistent, physically feasible, and reasoning-aligned by utilizing the reflection signal to modulate hidden activations with global semantic guidance.

In this work, we present CollabVLA, a self-reflective vision-language-action framework that transforms a standard visuomotor policy into a collaborative assistant. CollabVLA tackles key limitations of prior VLAs, including domain overfitting, non-interpretable reasoning, and the high latency of auxiliary generative models, by integrating VLM-based reflective reasoning with diffusion-based action generation under a mixture-of-experts design. Through a two-stage training recipe of action grounding and reflection tuning, it supports explicit self-reflection and proactively solicits human guidance when confronted with uncertainty or repeated failure. It cuts normalized Time by ~2x and Dream counts by ~4x vs. generative agents, achieving higher success rates, improved interpretability, and balanced low latency compared with existing methods. This work takes a pioneering step toward shifting VLAs from opaque controllers to genuinely assistive agents capable of reasoning, acting, and collaborating with humans.