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.
Figures & tables
Figure 1: The 2026 model. The VLM (left) encodes three camera images, register and scratch tokens, the task and stage embeddings and the robot state; the smaller action expert (right) denoises a 30-step action chunk and reads the VLM through a learnable mixture of the keys and values of all VLM layers. The task and stage embeddings also enter the expert directly through two conditioning tokens. Every attention layer of both transformers has a learnable sink. FAST tokens are used only in training.
Figure 2: Group-level attention mask in training (rows: queries, columns: keys; colored: can attend). Rows above the dashed line are VLM tokens, rows below are action-expert tokens. FAST tokens are causal among themselves and are removed from the cache the expert reads. The sink column is visible to every query in every layer.
Figure 3: Semantic stages (colors) of 24 demonstrations of make_microwave_popcorn against the five equal-time stages used in 2025 (dashed). Each row is one demonstration; the white gap at the end is the idle tail after task completion.
Figure 4: Raw attention of the action tokens over the three cameras (left wrist, right wrist, head). Left: the warm-started 2025 model on turning_on_radio ; the same lattice of cells lights up in all cameras and frames, regardless of content. Right: the 2026 model at step 60k, after inserting registers, scratch tokens and sinks, on make_microwave_popcorn ; the attention follows the popcorn bag that the robot reaches for. The color scales differ (left: 98th percentile of image tokens, right: absolute, saturating at 4 × uniform).
Figure 5: Flow-matching times of K=10 samples for four observations: independent draws (top) and stratified draws with antithetic noise pairs (bottom; gray lines are the quantile bins of Beta(1.5,1) , colors mark the ±ϵ pairs).
Figure 6: Left: gradient-variance budget per training step at a fixed batch, relative to the 2025 setting; only the flow term depends on the sampler and on K . Right: noise × step time as a function of K for i.i.d. and variance-reduced flow samples, from the measured costs and variance ratios; the shaded band is the flat region around the optimum.
Figure 7: The evaluation portal during a coffee-station rollout: per-instance q-scores (left), synchronized three-camera video and policy traces (center), and timestamped goal, grasp, and comment events (right).
Figure 8: Throughput of one-environment simulators sharing one RTX PRO 6000 (24 vCPUs). Six simulators exceed real time.
Figure 9: Observation and motion tools of the agent. Left: the three calls that Astra has, in practice and in evaluation. Middle: the controller executes one goal as one action per simulator step, and stops when the goal is reached, the motion stalls or the step limit passes. Right: the main iterations on the tools, each with the observation that confirmed it.
Figure 10: The Astra policy. Practice (left) uses training instances, resets, saved states and goal-predicate feedback to produce a frozen session, notes and recipes. Each agent-routed episode (right) forks that session. Astra plans goals from RGB-D and robot state; the controller uses robot-state feedback each simulator step and returns motion results, including stalls and step limits. Routing (bottom) selects Astra for 14 enabled tasks and the learned policy otherwise. Runtime errors or repeated query timeouts switch the rest of the episode to the learned policy; an unsuccessful motion alone does not trigger this fallback.
Figure 11: Public per-task Q, ordered by π0.5 score. On the 14 Astra tasks, stacked bars reach Astra’s Q; other bars show π0.5 Q.
Figure 12: Change in per-task Q on the 50 tasks shared with 2025, ordered by 2025 Q (lowest at left). Bars show 2026 public-evaluation Q minus 2025 submission Q; thin lines show 95% bootstrap intervals from ten episodes per task in each year. Missing 2025 episodes count as zero.
Figure 13: Manually annotated failure reasons on ten tasks shared by the 2025 submission and the 2026 learned policy. When an episode has multiple reasons, its count is divided equally among them.
Figure 14: Recovery and adaptation in the 2026 learned policy.
Vision-Language-Action (VLA) models offer a promising paradigm for general-purpose household robotics. However, inference and data-collection strategies that have been deployed in smaller-scale settings do not necessarily generalize to large-scale, diverse task suites. To address this gap, we conduct a systematic, paired evaluation of previously proposed inference and data-collection strategies across all 100 tasks. Our evaluation uses a fine-tuned π0.5 model based on the 2025 BEHAVIOR Challenge winner, which achieves a Q-score of 0.39 on the public test set. These results show that proposed strategies do not transfer uniformly to new environments and tasks. Project page, videos, and detailed analysis: https://rubenhammele.github.io/behavior2026
Ruben Hammele, Samira Huber, Mohammad Nazeri +7
VCAI Lab, Kiel University · RobotiXX Lab, George Mason University · AIR Lab, Ewha Womans University
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%.
Reliable execution of long-horizon mobile manipulation tasks remains challenging because overall task success depends on the successful completion of multiple constituent skills. Existing benchmarks, however, still rely primarily on full-task rollouts and aggregate task-level metrics, making intermediate failures difficult to observe and analyze. We present Behavior-Skill, a benchmark that reformulates the learning and evaluation of long-horizon tasks around executable constituent skills. It contains 235,492 skill instances from 10,000 demonstrations across 50 household tasks and 34 semantic skill categories. Each instance pairs a skill instruction with an aligned observation-action segment, and is further associated with a restorable intermediate state and a skill success condition to enable independent evaluation under valid preconditions. We further introduce trajectory-level and skill-level metrics to characterize policy capability beyond aggregate task success. Extensive experiments across representative VLA policies including pi0.5 and GR00T on the complete 50-task benchmark show that failures are highly non-uniform across skills, with contact-rich manipulation skills forming persistent bottlenecks. These results demonstrate that Behavior-Skill complements full-task evaluation by exposing intermediate capability profiles for analyzing and improving long-horizon VLA policies. Behavior-Skill is publicly available at https://github.com/nubot-nudt/Behavior-Skill.
Chunyun Ma, Lun Luo, Xingjian Luo +7
College of Intelligence Science and Technology and National Key Laboratory of Equipment State Sensing and Smart Support, National University of Defense Technology, Changsha, China. · XPeng Inc., Guangzhou, China. · The Chinese University of Hong Kong, Hong Kong, China. +1