This work presents the 2nd-place solution (Team IF1001) to the 2026 BEHAVIOR Challenge, which evaluates mobile manipulation across 100 household tasks. Our method is built on a powerful world action model (WAM), Cosmos 3. We improve Cosmos 3 by using richer visual history, task-progress conditioning, and task-group and skill-specific LoRA adapters. During inference, we apply a separate progress estimator to supply task-progress context, and use task- and progress-conditioned routing to select a task-group or skill-specific policy, complemented by action-candidate aggregation, rolling inpainting, and task-specific temporal and gripper control. To meet the minimum inference requirement of 1 FPS, we use neither recursive self-improvement (RSI) nor agentic harness methods and rely on no closed-source models; during official private-test evaluation, our system handled concurrent requests across 100 evaluation ports with limited hardware resources, sustaining an average end-to-end inference rate above 2 FPS per port. Our method achieves a mean Q-score of 0.613 and a full-task success rate of 40.9%, averaged over 3 self-evaluation rounds across all 100 tasks. On the official private test set, with 20 instances per task, our system achieves a mean Q-score of 0.615 and a full-task success rate of 41.2%.
Figures & tables
Characteristic
2025
2026
Household tasks
50
100
Human demonstrations per task
200
200
Mean human demonstration duration (s)
397.04
351.54
Distinct house-scale scenes
3
7
Skill annotation categories a
34
35
Public instances per task used for reporting
10
10
Table 1: Dataset scale and evaluation protocols of the 2025 and 2026 BEHAVIOR Challenges.
Figure 1: Overall framework of our method. The progress estimator predicts the subtask and stage ID from visual and state histories and the task ID. These predictions guide skill-specific LoRA activation, overriding the task-configured base policy or task-group LoRA adapter. Cosmos 3 is conditioned on visual history, task and skill text, current state, and the 4 conditioning tokens produced by stage-task fusion. Generated action chunks undergo candidate aggregation, retiming, and gripper control for closed-loop execution.
Figure 2: Attention mask. Following Cosmos 3 [ 5 ] , AR attention is causal within AR, while DM attends to both subsequences. The model jointly denoises future vision and action tokens.
Figure 3: Multimodal position encoding. (a) Packed token groups and representative coordinates for our policy, with the current observation at τ=2 s. Group sizes are schematic; visual tokens retain patch coordinates (hp,wp) , and dashed outlines denote prediction targets. Token order does not imply temporal order: current proprioception remains aligned with the current image. (b) Elapsed time maps to temporal-coordinate increments at 6 units per second.
Group
Tasks
Group
Tasks
1
1, 4, 37, 44, 83, 85
10
14, 31, 45, 52, 79, 89
2
3, 5, 10, 59, 68, 86
11
16, 32, 46, 61, 87, 92
3
0, 7, 8, 11, 69, 90
12
33, 47, 57, 70, 88, 93
4
2, 9, 12, 21, 25, 72
13
6, 23, 48, 58, 65, 94
5
17, 22, 26, 34, 54, 81
14
15, 50, 53, 71, 75, 97
6
20, 28, 30, 40, 74, 91
15
36, 55, 62, 76, 77, 98
Table 2: Task IDs assigned to each group.
LoRA adapter
Skill category
Pick
pick up from
Place
place in, place in next to, place on, place on next to, place under
Close
close door, close lid, close drawer
Open
open door, open lid, open drawer
Table 3: Skill-specific LoRA adapters and corresponding skill categories.
Figure 4: Progress estimator for stage conditioning and policy routing. The stage-estimation and subtask-prediction branches share a visual-history cache and task ID. Skill predictions and navigation/operation status guide the router; the dashed skill-to-stage lookup supplies alternative stage IDs for selected tasks.
Figure 5: Stage-ID sensitivity on task 26. Each cell shows the percentage change in mean action-prediction MAE relative to using the ground-truth stage ID on the same windows. Rows group windows by ground-truth stage; columns specify the input stage. Positive values indicate higher error.
Team / configuration
Method
Task IDs
Timeout
Val
Val
Test
Test
Q-score
SR (%)
Q-score
SR (%)
RLC (2025, 1st) [ 4 ]
VLA
0–49
2.0×Tˉtask
0.261
11.2
0.260
12.4
Openpi Comet (2025, 2nd) [ 3 ]
VLA
0–49
2.0×Tˉtask
0.183 a
14.4
0.251
11.4
Galaxea G0.5 b [ 21 ]
VLA
0–49
–
0.314
–
–
–
The North Star (2026, 1st)
Hiearchical Phys. RSI
0–99
1.5×Tˉtask
–
–
0.683
48.2
IF1001 (Ours, 2nd)
WAM
0–99
1.5×Tˉtask
0.613±0.008 c
40.9±1.0
0.615
41.2
Table 4: Overall, task-subset, and cherry-pick results. Official 2026 Test results and method names are taken from the official leaderboard . The 2025 and 2026 results follow their respective evaluation protocols (Table 1 ).
Configuration
No. of tasks
Val
Val
Q-score
SR (%)
Base policy
100
0.434
21.3
Base policy + task-group LoRA adapters
100
0.494
26.9
Table 5: Base policy model with and without task-group LoRA adapters.
Stage-ID strategy
No. of tasks
Val
Val
Q-score
SR (%)
Uniform episode clock
100
0.480
25.2
Stage estimation + post-processing
100
0.494
26.9
Table 6: Predicted versus clock-derived stage IDs for task-group LoRA adapters.
Configuration
No. of tasks
Val
Val
Q-score
SR (%)
Stage estimation + post-processing
24
0.681
52.5
Skill-to-stage lookup
24
0.707
56.3
Table 7: Skill-to-stage lookup on 24 tasks with complete results in both configurations.
Configuration
Model
No. of tasks
Val
Val
calls (k)
Q-score
SR (%)
Short-window
676.7
100
0.575
35.9
Long-window
340.0
100
0.549
33.5
TOPP-RA
540.3
100
0.557
36.3
Table 8: Task performance and policy-call cost of execution-window and retiming strategies.
Configuration
Mean time (s)
Human demonstrations
313.12
Uniform temporal resampling
274.57
TOPP-RA
211.40
Table 9: Task-averaged execution time across 60 tasks. Robot results use 260 instances successful under both configurations; the human reference uses the same tasks.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 6: Initial observations for the 17 task groups.
Figure 7: Head-camera views during demonstrated skill execution: (a) picking up a radio, (b) placing a cookie in a basket, (c) opening a microwave door, and (d) closing a washing-machine door.
Task ID
Task
Val: Round 0
Val: Round 1
Val: Round 2
Q-score
SR (%)
Q-score
SR (%)
Q-score
SR (%)
0
Turning on radio
0.700
70.0
0.700
70.0
0.600
60.0
1
Picking up trash
1.000
100.0
1.000
100.0
1.000
100.0
2
Putting away Halloween decorations
0.800
40.0
0.886
60.0
0.800
40.0
3
Cleaning up plates and food
0.586
40.0
0.643
50.0
0.500
30.0
4
Can meat
0.211
0.0
0.356
20.0
0.322
10.0
Appendix
Table 10: Per-task validation results across 3 evaluation rounds (100 tasks, 10 instances per task per round).
Group
No. of tasks
Base policy
Task-group LoRA
Val
Val
Val
Val
Q-score
SR (%)
Q-score
SR (%)
1
6
0.485
38.3
0.509
38.3
2
6
0.507
20.0
0.559
21.7
3
6
0.380
20.0
0.351
18.3
4
6
0.575
15.0
0.626
30.0
Appendix
Table 11: Base policy versus task-group LoRA fine-tuning within each task group.
Figure 8: Subtask–stage associations in demonstration data: (a) concentrated stage distributions in task 40 and (b) dispersed distributions in task 54. Each row shows the percentage of samples for a given subtask occurring at each stage. Darker cells indicate larger percentages, and outlined cells identify the stage selected by lookup. The mean row maximum summarizes concentration across subtasks. Row totals may differ from 100% due to rounding.
World action models (WAMs) built on video generation backbones are a rising recipe for robot learning, yet remain confined to tabletop manipulation. Mobile manipulation demands simultaneous locomotion and whole-body manipulation amid scene-scale dynamics, yet is still dominated by dynamics-blind visual encoders with hand-crafted coordination. We bridge this gap with MobileWAM, a mixture-of-transformers architecture that fuses a pretrained video diffusion transformer with a lightweight action expert through layerwise joint attention, translating internet-scale motion priors into whole-body control. To reconcile the heterogeneous dynamics of moving and manipulating, each feed-forward layer of the action expert becomes a three-expert mixture of shared, locomotion, and manipulation experts, softly routed by the motion intent in the action tokens. To densify supervision, we further propose Chain-of-Foresight (CoF): intermediate representations sequentially predict a chain of future latent chunks, each step conditioned on its predecessor. CoF pairs naturally with our decoupled video--action denoising scheme. At deployment, the WAM serves as a pure current-frame encoder; foresight acts only through gradients, so at inference the foresight chain and video generation are discarded, leaving only policy-level cost. MobileWAM surpasses state-of-the-art mobile manipulation policies on ManiSkill-HAB and fine-tunes to a real ARX Lift2 mobile manipulator across diverse tasks with strong generalization. Code will be released soon.
Zehua Fan, Junjie He, Wenxuan Song +14
Institute for AI Industry Research (AIR), Tsinghua University · 2Shanghai Jiao Tong University · 3The Hong Kong University of Science and Technology (Guangzhou) +8
We describe our solution for the 2026 BEHAVIOR Challenge: 100 long-horizon household tasks in the OmniGibson simulator. We placed 9th on the final leaderboard, the only independent team in the top 10. The core of the solution is a single vision-language-action (VLA) model, one checkpoint for all 100 tasks, built on our 1st-place 2025 solution, which is based on π0.5. On the 50 tasks shared with 2025, its Q rises from 0.260 to 0.371; across all 100 tasks it scores 0.392. Our main conclusion from 2025 was that the model was undertrained, so most of the 2026 work went into training the model for longer, optimizing some training parameters, making each training step slightly faster and reducing the gradient noise: a migration from 50 to 100 tasks (semantic stages built from the demonstrations' skill annotations, task and stage embeddings of new tasks warm-started from similar old ones, and a 21-dimensional action space matching the teleoperation controls); variance-reduced multi-sample flow matching with a simple noise-vs-compute model that picks the number of action-expert samples per observation without ablations; and a 200k-step training run, about one pass over the 211M frames of the demonstrations. A cloud evaluation infrastructure keeps a shared record with provenance, episode diagnostics and restorable simulator states, and runs six simulators on one GPU together faster than real time. On top of the learned policy, 14 tasks are routed to a GPT-6 Astra coding agent that controls the robot through observation and motion tools, starting every episode from a frozen copy of its own practice session on training instances. Across these 14 tasks, Astra's mean Q is 0.639, against 0.108 for the learned policy. The public submission included only part of the Astra results and scored 0.446 Q with 27.3% full-task success.
Mobile manipulation is a key capability for general-purpose robots, yet remains challenging for current embodied learning methods. VLA policies are typically reactive and lack explicit world modeling, while existing World Action Models (WAMs) are still poorly aligned with the structure of mobile manipulation: they operate on coarse video chunks, model entangled navigation-manipulation actions, and train inverse dynamics under supervision that does not match autoregressive inference. As a result, they often miss fine-grained contact dynamics, suffer from action-distribution conflicts, and accumulate errors over long-horizon rollouts. We propose ABot-M0.5, a new WAM built on the insight that mobile manipulation requires alignment at three levels: temporal granularity, action space, and train-test consistency. To align temporal granularity, we introduce intermediate latent actions that capture local visual state transitions and serve as an bridging action space between video latents and embodiment-specific controls. To align action space, we design a dual-level Mixture-of-Transformers architecture that disentangles both modality representations and heterogeneous action subspaces such as base movement and arm manipulation. To align inference conditions, we propose the dream-forcing training strategy that progressively trains inverse dynamics on model-predicted videos, improving train-test alignment and robustness during autoregressive prediction. Experiments on challenging mobile and fine-grained manipulation benchmarks demonstrate that ABot-M0.5 achieves state-of-the-art performance in both long-horizon task success and finegrained control accuracy. These results highlight the critical importance of granularity-aligned, action-disentangled, and inference-consistent world-action modeling.