BICPO-VLA: Behavior-Identified Continuation Preference Optimization for Smooth Asynchronous Vision-Language-Action Control
Authors: Ming Shang, Yuchen Huang, Jiaoyang Chen, Haoyuan Hu, Han Yu, Liping Song, Luyun Feng, Shuo Bao, Wei Dong, Xinzhou Wang, Fuchun Sun
Organizations: Beihang University, Beijing, China · Beijing Institute of Technology, Beijing, China · Institute of Automation, Chinese Academy of Sciences, Beijing, China · Beijing University of Posts and Telecommunications, Beijing, China · Hunan University, Changsha, China · Peking University, Beijing, China · Tsinghua University, Beijing, China
The request-to-handoff gap has three coupled sources: ambiguity about the behavior intended at request time, physical-state drift accumulated during action generation, and residual incompatibility when the new action finally assumes control. BICPO-VLA addresses them in sequence. First, an instruction-aware causal history encoder identifies the behavior supported by the command and current task progress. Second, sequential Haar subspace generation decomposes each action chunk into complementary pairwise scaffold and residual coefficients, enabling two specialized generation stages followed by exact reconstruction. By reducing iterative refinement in the original action space, it shortens the interval over which the robot continues moving before the new chunk becomes available. Finally, BICPO rolls the known outgoing actions to the actual handoff state and applies reference-relative Flow-DPO among behaviorally matched candidates, adapting the generated chunk to the remaining request-to-handoff mismatch without changing its intended behavior.
Figures & tables
Figure 1: Motivation and fibered view of asynchronous behavior-to-action realization. Left: representative VLA paradigms improve memory, reduce generative inference steps, or regularize generic smoothness, while retaining a largely direct mapping from multimodal context to an action chunk. Center: BICPO-VLA makes behavior identity explicit through instruction-aware task-progress reasoning and behavior memory. Right: the instruction identifies a behavior-conditioned action fiber; Haar coordinates organize its realizations into a pairwise motion scaffold and a complementary local residual; and the rolled handoff state with BICPO selects the realization that can take control coherently. Bottom: improvements across simulation and real-world settings.
Figure 2: BICPO-VLA as handoff-conditioned action realization. Left: instruction-aware visual–action encoding and behavior memory identify the behavior condition supported by the command and execution history. Center: a fixed Haar transform organizes realizations of that behavior into a pairwise motion scaffold and a complementary local residual, which are generated sequentially and reconstructed exactly. Right: the rolled handoff state specifies the execution context, and BICPO compares same-fiber candidates by boundary displacement and motion-trend mismatch. The reference pathway remains frozen, while only the continuation projection and selected flow-policy parameters are updated.
Method
1
2
3
4
5
Avg.
RoboVLM
98.0
93.6
85.4
77.8
70.4
4.25
ReconVLA
95.6
87.6
76.9
69.3
64.1
3.95
OpenVLA
91.3
77.8
62.0
52.1
43.5
3.27
UniVLA
95.5
85.8
75.4
66.9
56.5
3.80
UP-VLA
92.8
86.5
81.5
76.9
69.9
4.08
RoboDual
94.4
82.7
72.1
62.4
54.4
3.66
Table 1: Results on CALVIN ABC → D. Columns 1–5 report the percentages of evaluation chains completing at least the corresponding number of subtasks. Avg. denotes the average completed sequence length.
Figure 3: Qualitative simulation comparison. BICPO-VLA (top) and π0.5 (bottom) on two multi-stage tasks. BICPO-VLA preserves the intended phase through decisive grasp, placement, and button-activation transitions; circles mark terminal outcomes.
Task Group I
Task Group II
Summary
Method
AB
CA
CB
DB
GR
MP
P2R
PB
BF
CP
Avg. I
Overall
RDT
75
12
9
32
43
11
1
2
27
17
34.2
22.9
ACT
23
4
3
1
25
0
0
0
0
1
11.2
5.7
DP
0
5
0
0
0
0
0
0
0
0
1.0
0.5
DP3
3
14
3
53
2
3
0
1
18
1
15.0
9.8
π0
56
11
6
24
80
22
6
4
4
45
35.4
25.8
Table 2: Success rates (%) on the ten RoboTwin 2.0 Hard tasks. Task Group I contains AB, CA, CB, DB, and GR, with Avg. I reporting their mean success rate. Task Group II contains MP, P2R, PB, BF, and CP, while Overall reports the mean success rate across all ten tasks. The best and second-best results are highlighted in bold and underlined, respectively. Full task names are provided in the technical supplement.
Method
SR (%) ↑
Jjump↓
Jtrend↓
BICPO-VLA (Ours)
99.1
2.37
6.40
BICPO-VLA without DPO
98.8
3.01
8.07
BICPO-VLA with chosen-only SFT
97.9
2.85
6.80
π0.5 -FM
96.9
4.22
8.50
π0.5 -FM with chosen-only SFT
49.1
3.13
6.90
π0.5 -FM with continuity DPO
97.1
2.52
6.70
Table 3: LIBERO task success and handoff continuity. Chosen-only SFT imitates preferred continuations, while continuity DPO optimizes preferred over rejected continuations relative to a reference model.
Figure 4: Real-world evaluation. Top: BICPO-VLA and π0.5 rollouts on milk placement and paper-ball disposal. Bottom: success rates on six tasks; BICPO-VLA performs best throughout.
Metric
k=3
k=4
k=5
Random
SR (%) ↑
99.1
98.9
98.8
98.8
Jjump↓
2.37
2.47
2.50
2.46
Jtrend↓
6.40
6.21
6.23
6.17
Table 4: LIBERO robustness to fixed and random inference delays, with k∈{3,4,5} .
Variant
1
2
3
4
5
Avg.
Full
98.9
95.4
91.0
86.0
80.7
4.520
w/o Haar
98.2
94.8
90.7
85.2
79.6
4.490
w/o BICPO
97.8
95.0
90.5
85.4
78.9
4.480
w/o Instr.
97.0
92.5
87.9
84.0
78.2
4.396
Table 5: Component ablation on CALVIN ABC → D; w/o Haar predicts actions in the original coordinate space.
Variant
1
2
3
4
5
Avg.
Visual query
98.9
95.4
91.0
86.0
80.7
4.520
Query fusion
98.6
94.4
89.2
85.1
80.3
4.476
Memory fusion
97.5
93.1
88.3
84.2
78.6
4.417
Table 6: Instruction-intervention ablation on CALVIN ABC → D with matched capacity.