VQActFlow: Vector-Quantized Action Mode Steering for Multi-Task Robot Manipulation

Abstract

In multi-task manipulation, a policy has to commit to the action mode that matches the current instruction. Pick the wrong mode and the robot does the wrong task. VQActFlow tokenizes action chunks into a discrete codebook and generates the codes with Variational Flow Matching, so it carries a distribution over action modes through every step of generation. At inference time, classifier-free guidance and a learned codebook critic reshape that distribution. Neither one retrains the policy.

81.0% LIBERO-Goal peak success (w=4), vs. 61.5% for the closest discrete baseline

80.5% LIBERO-90 success with CFG and the critic, best of every baseline we ran

57.5% G1 humanoid success at w=6, up from 23.8% with no guidance

77.5% Bimanual average success with CFG and the critic together

Method

How VQActFlow represents and steers actions

VQActFlow framework: VQ-VAE action tokenization followed by a Variational Flow Matching policy with CFG and codebook critic guidance. — **Figure 1.** A VQ-VAE tokenizes action chunks into a discrete codebook. A Variational Flow Matching policy then generates code sequences, keeping logits over the full codebook at every step. CFG and a codebook critic act on those logits to steer the result.

Most flow-matching policies work in a continuous space with no explicit notion of action mode, so guidance can only reshape a single continuous distribution. VQActFlow instead carries a discrete distribution over action modes through generation, which gives guidance a categorical target to act on from the first step to the last.

1
CFG on language conditioning. At every step it extrapolates between the conditional and unconditional logits, which sharpens how firmly the policy commits to the instructed task.
2
A learned codebook critic. A small transformer, trained contrastively, scores how feasible each code is and pushes the distribution toward feasible modes.

Both operate on the same categorical distribution, so you can apply them together, and neither one needs extra training.

Experiments

A G1 humanoid, a bimanual platform, and LIBERO in simulation

Across all three platforms, VQActFlow beats both continuous and discrete baselines trained on the same encoder, data, and step budget.

Humanoid whole-body manipulation

A Unitree G1 with two dexterous hands runs four pick-and-place tasks in one shared workspace; only the target object and the instruction change.

Place the red ball into the box Success ×1.5

Figure 2. G1 humanoid whole-body pick-and-place. Pick a task to watch VQActFlow run it.

**Table I.** Aggregated G1 hardware policy success rate (%).
CFG weight	Success	Missed grasp	Wrong task
w=1	23.8	35.0	41.3
w=6	57.5	40.0	2.5

Almost all of CFG's gain here comes from fixing wrong-task errors, which fall from 41.3% to 2.5%. It helps the policy choose the right task; it does not make the motion itself any cleaner.

Bimanual manipulation

Two ALOHA-style arms on four contact-rich tasks, three of them needing coordinated bimanual use.

Sweep the battery into the dust pan Success ×1.5

Figure 3. The bimanual setup: two ALOHA-style arms, one wrist camera each, and a fixed workspace camera. Pick a task to view a rollout.

**Table II.** Success rates (%) on bimanual manipulation tasks.
Method	Battery	Duck	Cylinder	Ball	Avg.
Discrete Policy, w=1.0	35.0	70.0	15.0	75.0	48.8
Discrete Policy, w=4.0	65.0	70.0	35.0	75.0	61.3
VQActFlow, w=1.0, λ=0.0	40.0	70.0	15.0	55.0	45.0
VQActFlow, w=1.0, λ=1.0	50.0	75.0	50.0	65.0	60.0
VQActFlow, w=4.0, λ=0.0	70.0	85.0	55.0	85.0	73.8
VQActFlow, w=4.0, λ=1.0	70.0	90.0	65.0	85.0	77.5

With the same backbone, a continuous flow-matching baseline fails all four tasks. Its joints oscillate violently, roughly two orders of magnitude rougher than VQActFlow.

VQActFlow: success

CFM baseline: failure

Figure 4. The same task and the same backbone, shown in real time. VQActFlow completes the place; the continuous-flow CFM baseline oscillates and fails.

**Table III.** Bimanual inference time per action chunk (ms), RTX 5090.
Configuration	Unguided	CFG	Critic	CFG + critic
Inference time	155.2	288.9	216.7	350.2

350 ms fits comfortably inside the 1.28 s budget per action chunk, and asynchronous inference computes the next chunk while the current one runs.

CFG on LIBERO-Goal

Guiding the discrete distribution directly works better than guiding a continuous approximation of it.

LIBERO-Goal success rate vs. CFG weight for VQActFlow and Discrete Policy. — **Figure 5.** Success rate vs. CFG weight. VQActFlow peaks at w=4 (81.0%); Discrete Policy needs stronger guidance and still tops out lower (61.5%).

Multi-task scaling on LIBERO-90

At this scale, CFG and the codebook critic each contribute their own gains, and the gains mostly stack.

**Table IV.** LIBERO-90 success rate (%) for different policies, trained from scratch under matched conditions.
Method	Type	CFG w	Critic λ	Success
MT-ACT	Continuous	–	–	72.4
CFM	Continuous	–	–	77.2
VQ-BeT	VQ-based	–	–	24.1
Discrete Policy	VQ-based	1.0	–	49.4
Discrete Policy	VQ-based	2.0	–	60.3
VQActFlow	VQ-based	1.0	0.0	72.3
VQActFlow	VQ-based	1.0	1.0	74.3
VQActFlow	VQ-based	2.0	0.0	77.6
VQActFlow	VQ-based	2.0	1.0	80.5

**Table V.** Ablation on codebook size K, LIBERO-90, guidance disabled.
Codebook size	128	256	512	1024
Success rate	62.6	65.6	72.3	70.1

Success peaks at K=512. Smaller codebooks limit how finely actions can be quantized; larger ones blow up the per-position classification space and make the categorical prediction harder to learn.

Citation

BibTeX

@article{anonymous2026vqactflow,
  title={VQActFlow: Vector-Quantized Action Mode Steering for Multi-Task Robot Manipulation},
  author={Anonymous Author(s)},
  journal={Under review},
  year={2026}
}