FATE: Closed-Loop Feasibility-Aware Task Generation with Active Repair for Physically Grounded Robotic Curricula

Bingchuan Wei; Bingqi Huang; Jingheng Ma; Zeyu zhang; Sen Cui

cs.ROMar 2, 2026

FATE: Closed-Loop Feasibility-Aware Task Generation with Active Repair for Physically Grounded Robotic Curricula

Authors: Bingchuan Wei, Bingqi Huang, Jingheng Ma, Zeyu zhang, Sen Cui

Organizations: School of Aerospace Engineering, Tsinghua University, Beijing, China · School of Integrated Circuits, Tsinghua University, Beijing, China · State Key Laboratory of General Artificial Intelligence, Beijing Institute for General Artificial Intelligence (BIGAI), Beijing, China · Department of Automation, Tsinghua University, Beijing, China

Abstract

Recent breakthroughs in generative simulation have harnessed Large Language Models (LLMs) to generate diverse robotic task curricula, yet these open-loop paradigms frequently produce linguistically coherent but physically infeasible goals, stemming from ungrounded task specifications or misaligned objective formulations. To address this critical limitation, we propose FATE (Feasibility-Aware Task gEneration), a closed-loop, self-correcting framework that reimagines task generation as an iterative validation-and-refinement process. Unlike conventional methods that decouple generation and verification into discrete stages, FATE embeds a generalist embodied agent directly into the generation loop to proactively guarantee the physical groundedness of the resulting curriculum. FATE instantiates a sequential auditing pipeline: it first validates static scene attributes (e.g., object affordances, layout compatibility) and subsequently verifies execution feasibility via simulated embodied interaction. Critical to its performance, upon detecting an infeasible task, FATE deploys an active repair module that autonomously adapts scene configurations or policy specifications, converting unworkable proposals into physically valid task instances. Extensive experiments validate that FATE generates semantically diverse, physically grounded task curricula while achieving a substantial reduction in execution failure rates relative to state-of-the-art generative baselines.

Figures & tables

Figure 1: FATE aligns task feasibility across static and dynamic dimensions. (Left) Static Alignment: The system corrects initial scene priors by resolving semantic violations (e.g., replacing raw eggs), geometric chaos, and precondition conflicts. (Right) Dynamic Alignment: FATE addresses runtime execution failures, such as high-torque grasps, unreachable targets, and collisions, through iterative rollback mechanisms. Red highlights indicate initial infeasibilities; Green indicates FATE-repaired executable environments.

Figure 2: Visualization of the Hierarchical Feasibility Alignment Process. The optimization trajectory evolves in the joint task space $\tau=(\mathcal{I},\mathcal{S},\Pi)$ . (Left) Static Alignment ( $\mathcal{A}_{perc}$ ): The Ante-Auditor corrects Scene ( $\mathcal{S}$ ) and Instruction ( $\mathcal{I}$ ) incompatibilities via discrete repairs (e.g., RESCALE ), projecting $\tau_{init}$ onto the perceptually valid manifold. (Right) Dynamic Alignment ( $\mathcal{A}_{exec}$ ): Targeting the Policy ( $\Pi$ ) , the system refines the solver configuration using semantic feedback and dynamic APIs (e.g., INJECT REWARD ), guiding the task from failure (red trajectories) to the Feasible Space (green polyhedron).

Figure 3: Evolutionary Trajectory of the ”Heat Food” Task. Starting from a chaotic initialization ( $\tau_{init}$ ), FATE applies Static Repair to rearrange objects and correct affordances (swapping an egg for a meal). During execution, the system autonomously detects and repairs dynamic failures: a grasp failure is resolved via Motion Re-planning , and an insertion collision is fixed via Asset Resizing . The gray curved arrows illustrate our Sub-step Rollback mechanism, which invokes specific repair APIs (bottom icons) to iteratively refine the task into a successful policy.

Semantic Diversity (Language)
	RLBench	ManiSkill2	Meta-World	Behavior-100	GenSim-V2	FATE (Ours)
Number of Tasks	106	20	50	100	70	352
Task Desc. - Self-BLEU-4 ( $\downarrow$ )	0.317	0.674	0.322	0.299	0.378	0.276
Task Desc. - S-BERT Sim ( $\downarrow$ )	0.200	0.194	0.263	0.210	0.288	0.158
Visual & Configural Diversity (Scene)
Scene Image - ViT Sim ( $\downarrow$ )	0.375	0.332	0.517	0.389	0.717	0.189

Table 1: Generative Diversity Analysis. We compare the semantic and visual diversity of tasks generated by FATE against expert-designed benchmarks and open-loop baselines. FATE achieves diversity comparable to human-curated benchmarks while ensuring physical feasibility.

Baselines
	Scene Validity (SVR)	Execution Validity (EVR)	Feasible Task Rate (FTR)
	(Static Validity) ( $\uparrow$ )	(Dynamic Validity) ( $\uparrow$ )	(End-to-End Yield) ( $\uparrow$ )
Vanilla (No Audit)	41.5%	30.4%	12.6%
GenSim-V2	58.2%	51.2%	29.8%
FATE (Ours)
FATE-Static (Ante-Audit)	93.8%	69.7%	65.4%

Table 2: Quantifying the Feasibility Gap. We compare the validity of tasks generated by the vanilla baseline against FATE. We decompose the total yield (FTR) into static Scene Validity (SVR) and dynamic Execution Validity (EVR).

Diagnostic Category	Precision	Recall	F1-Score
Semantic	0.96	0.94	0.95
Geometric	0.88	0.85	0.86
Dynamic	0.85	0.81	0.83
Overall	0.90	0.87	0.88

Table 3: Diagnostic Performance of the Auditor. We report the Precision, Recall, and F1-Score of the FATE Auditor across three error categories on a human-labeled test set.

Repair Stage	Detected Failures (Input)	Successful Repairs (Output)	RSR (%)
Ante-Repair (Static)	215	201	93.5%
Primitive-Repair (Runtime)	42	39	92.9%
RL-Repair (Learning)	76	58	76.3%
Aggregate	333	298	89.5%

Table 4: Repair Efficacy Analysis. We report the Repair Success Rate (RSR) across three stages. The ”Detected Failures” represents tasks identified as lying outside the manifold, and ”Successful Repairs” denotes tasks successfully projected onto the manifold.

Figure 4: Task-Wise Success Rate Analysis. We evaluate performance across five tasks to disentangle the contributions of the alignment operators. The Static Alignment ( $\mathcal{A}_{perc}$ ) module excels in geometry-sensitive tasks (e.g., Ingredient into Pot ), while the Dynamic Alignment ( $\mathcal{A}_{exec}$ ) module is critical for contact-rich tasks (e.g., Sliding Window ). FATE (Green) combines both to achieve robust performance.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

API Signature	Functionality & Theoretical Mapping
SWAP_ASSET(tgt_id, query)	Substitutes an object to maximize Affordance Matching . Used when $\Psi_{\mathcal{I}}$ cannot be satisfied by the current object semantics.
SPAWN_ASSET(query, pose)	Instantiates missing nodes in the Scene Graph $\mathcal{S}$ . Used when the instruction $\mathcal{I}$ implies objects not present in $\mathcal{S}_{init}$ .
REMOVE_ASSET(tgt_id)	Prunes the Scene Graph. Used to remove clutter that violates Geometric Reachability or creates collision conflicts.
SET_JOINT(tgt_id, idx, val)	Forces configuration state $q_{j}$ to a specific value. Used to resolve logical inconsistencies (e.g., ”Open a door” that is already open).
TRANSFORM_POSE(tgt_id, $\Delta$ )	Applies a rigid body transform. Used to resolve interpenetration ( Physical Plausibility ).
RESCALE(tgt_id, factor)	Scales the physical manifold $M_{j}$ . Directly addresses Morphological Compatibility (e.g., resizing an object to fit a container).

Appendix

Table 5: Static Repair API (

\mathcal{A}_{perc}

). These primitives implement the discrete update

\Delta_{static}

to resolve perceptual hallucinations defined in Section 4.3 .

API Signature	Functionality & Theoretical Mapping
RL-Specific Optimizations (Policy Manifold Shaping)
INJECT_REWARD(func_str)	Overwrites the reward function with code synthesized by the Brain. This implements Reward Landscape Synthesis , allowing the VLM to reshape the optimization surface to guide exploration toward $\Phi_{\tau}$ .
SET_RL_HORIZON(steps)	Adjusts the maximum episode length. This expands the temporal search space of $\Omega_{\pi}$ , ensuring sufficient time for convergence.
Execution Dynamics & Planner Tuning
REPLAN_PATH(inflation_r)	Modifies configuration space constraints for the MPC solver. Used to resolve collision-induced deadlocks.
ADJUST_GRASP(offset, vector)	Modifies the end-effector approach pose. Used to correct high-torque grasp failures or singularities.

Appendix

Table 6: Dynamic Repair API (

\mathcal{A}_{exec}

). These primitives instantiate the semantic policy gradient

\nabla_{sem}

, modifying the solver landscape to capture feasible solutions.

Synthetic Data Generation for Robotics

Synthetic data generation for robotics creates simulated scenes, sensor readings, and action trajectories to train and evaluate robots.

See all 301 papers

RoboGene: Diversity-Driven Agentic Task Generation Framework for VLA Pre-training
Feb 18, 2026Yixue Zhang, Kun Wu, Zhi Gao +12Synthetic Data Generation for Robotics Robotic Data Collection
Beijing Innovation Center of Humanoid Robotics · The School of Advanced Manufacturing and Robotics, Peking University · Beijing Institute of Technology +2
Generative Flow on Symbolic Abstractions for Embodiment Capability Discovery
Oct 12, 2026Youwei Yu, Lantao LiuLong-Horizon Robotic Manipulation
Indiana University Bloomington
Learning Actionable Manipulation Recovery via Counterfactual Failure Synthesis
Mar 13, 2026Dayou Li, Jiuzhou Lei, Hao Wang +6Robot Failure Recovery Synthetic Data Generation for Robotics
Texas A&M University · University of Minnesota · University of Texas at Austin +2

Feb 18, 2026J/K move · Enter open · S save

RoboGene: Diversity-Driven Agentic Task Generation Framework for VLA Pre-training

Yixue Zhang, Kun Wu, Zhi Gao, Zhen Zhao+11

Beijing Innovation Center of Humanoid Robotics · The School of Advanced Manufacturing and Robotics, Peking University · Beijing Institute of Technology+2

Synthetic Data Generation for Robotics Robotic Data Collection Vision-Language-Action Models Generalization in Robotic Manipulation

Open paper AskPDF arXiv

The pursuit of general-purpose robotic manipulation is hindered by the scarcity of diverse, real-world interaction data. Unlike large-scale web vision-language data, robotic data collection incurs prohibitive physical costs. The key challenge is thus not merely to collect more robot data, but to obtain informative training data that broadens behavioral coverage across objects, scenes, and skills. Yet automatically curating task proposals to maximize the utility of collected data remains under-explored. We introduce RoboGene, an agentic framework for generating diverse, informative manipulation tasks across single-arm, dual-arm, and mobile robots. RoboGene combines (1) diversity-driven sampling for broad task coverage, (2) agentic self-reflection for task-quality assessment, and (3) human-verified feedback memory for execution-grounded validation and continual refinement. Through quantitative analysis and large-scale real-world experiments, we collect 18k trajectories and introduce metrics for task feasibility and diversity. RoboGene significantly outperforms frontier foundation models, including GPT-4o and Gemini 2.5 Pro. VLA models trained on RoboGene-generated data achieve higher success rates and stronger generalization, highlighting the value of high-quality task generation. Project page: https://robogene-boost-vla.github.io/. Datasets: Hugging Face (https://huggingface.co/datasets/X-Humanoid/RoboGene) and ModelScope (https://modelscope.cn/datasets/X-Humanoid/RoboGene_tasks).

Abstract

Synthetic Data Generation for Robotics

Explore similar work

RoboGene: Diversity-Driven Agentic Task Generation Framework for VLA Pre-training

Generative Flow on Symbolic Abstractions for Embodiment Capability Discovery

Learning Actionable Manipulation Recovery via Counterfactual Failure Synthesis