Neurosymbolic Embodied Agents

Mohammad Albinhassan; Yuming Feng; Alessandra Russo; Pranava Madhyastha

cs.ROAug 17, 2026

Neurosymbolic Embodied Agents

Authors: Mohammad Albinhassan, Yuming Feng, Alessandra Russo, Pranava Madhyastha

Organizations: Department of Computing Imperial College London · Department of Computer Science Johns Hopkins University, Whiting School of Engineering · Department of Computer Science City, University of London

Abstract

Language and vision-language models generate plausible embodied plans but do not guarantee executability, as their outputs can violate environment dynamics or act on incorrectly grounded entities. We present a neurosymbolic agent that factors long-horizon household tasks into task-directed visual exploration and constrained symbolic planning. In the first phase, a vision-language model and exploration harness acquire goal-relevant predicates and instance bindings from egocentric observations and grounded interactions, producing a symbolic initial state. In the second, a PDDL transition model restricts decoding to tokens that extend applicable actions. Monte Carlo tree search then evaluates executable continuations using a domain-independent planning heuristic. The resulting plans are executable by construction under the transition model, with transfer to the environment conditioned on correct visual grounding. On VirtualHome and ALFWorld, open 4B-27B models exceed 90% success in both environments, and our smallest agent substantially outperforms a 27B direct visual policy in each. Constraints and search prove complementary rather than interchangeable: in ALFWorld either alone solves under a third of tasks, whereas their combination solves over 95%. The method also uses several times fewer generated tokens than extended thinking and far fewer model-visible images than direct interaction, and residual failures localize to state acquisition rather than plan generation without any specialized training.

Figures & tables

Panel of Figure 1 — Figure 1: Two phases joined by a narrow symbolic interface. A literal enters $\mathcal{F}_{t}$ only under visual or interaction evidence, so priors decide where to look , not what is present (greyed chip: unsupported). Only $\widehat{s}_{0}$ and $\beta_{\tau}$ cross into Phase II, which excludes at three scales: inapplicable actions (struck pill), illegal tokens (slashed cells), unexpanded branches (struck node).

Neurosymbolic embodied agent
Approach	Model	VH	ALFW
Ours	Qwen3.5-4B	94.5	90.3
Ours	Qwen3.5-9B	99.5	97.8
Ours	Qwen3.5-27B	99.0	95.5
Direct Qwen policies
Direct VLM	Qwen3.5-4B	5.5	1.5

Table 1: Task success (%) on VirtualHome and ALFWorld. Direct VLM and embodied VLM policies predict one environment action per interaction step. Best results are bold.

Figure 2: Inference and environment-observation cost across model scales. (a) Mean generated tokens per task for unconstrained thinking, unconstrained MCTS, and our constrained MCTS. (b) Mean model-visible images per task for direct interactive policies and our two-phase agent. Lower is better in both panels.

Figure 3: Terminal failure attribution, macro-averaged equally across VirtualHome and ALFWorld. Values are percentages of all tasks.

Method	Search	Constraints	VH	ALFW
Greedy	–	–	61.0	15.4
Thinking	–	–	59.0	16.9
Greedy	–	✓	98.5	32.3
MCTS	✓	–	99.0	29.2
Ours	✓	✓	99.0	95.5

Table 2: Component ablation with Qwen3.5-27B. Entries report end-to-end task success (%).

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Condition	Thinking	Temperature	Top- $p$	Top- $k$	Max output tokens
Qwen Phase I exploration	off	0.0	1.0	–	512
Qwen direct interactive	off	0.0	1.0	–	128 per turn
Qwen greedy Phase II	off	0.0	1.0	–	512
Qwen MCTS Phase II	off	0.0	1.0	–	128 per expansion
Qwen thinking, main paper	on	0.0	1.0	–	32,768
Qwen thinking, supplementary	on	1.0	0.95	20	32,768

Appendix

Table 3: Decoding settings. A dash denotes no top-

k

cutoff. Qwen supplementary thinking, MiMo, Embodied-R1.5, and RoboBrain nucleus sampling are evaluated with seeds 0, 42, and 1738.

Benchmark	Per episode	One run	Three runs
VirtualHome	$0.0148	$2.97	$8.90
ALFWorld	$0.0534	$7.16	$21.48
Combined	–	$10.13	$30.38

Appendix

Table 4: Estimated Gemini 3 Flash high-thinking API cost in USD.

		VH		ALFW
Approach	Model	Accuracy (%)	95% CI	Accuracy (%)	95% CI
Ours	Qwen3.5-4B	94.5 $\pm$ 0.0	[91.0, 97.5]	90.3 $\pm$ 0.0	[85.1, 94.8]
Ours	Qwen3.5-9B	99.5 $\pm$ 0.0	[98.5, 100.0]	97.8 $\pm$ 0.0	[94.8, 100.0]
Ours	Qwen3.5-27B	99.0 $\pm$ 0.0	[97.5, 100.0]	95.5 $\pm$ 0.0	[91.8, 98.5]
Direct VLM	Qwen3.5-4B	5.5 $\pm$ 0.0	[2.5, 9.0]	1.5 $\pm$ 0.0	[0.0, 3.7]
Direct VLM	Qwen3.5-9B	10.0 $\pm$ 0.0	[6.0, 14.5]	6.0 $\pm$ 0.0	[2.2, 10.4]

Appendix

Table 5: Accuracy and robustness across model conditions. The value after

\pm

is the sample standard deviation across three runs. Deterministic conditions are marked

\pm 0.0

. Intervals are 95% percentile task-cluster bootstrap intervals with repeated-run outcomes held together within each resample. Bold denotes the highest accuracy in each benchmark and results not significantly different from it.

	VH		ALFW
Comparison	Holm $p$	Sig.	Holm $p$	Sig.
Ours (Qwen3.5-4B) vs. Direct VLM (Qwen3.5-4B)	$<0.001$	✓	$<0.001$	✓
Ours (Qwen3.5-9B) vs. Direct VLM (Qwen3.5-9B)	$<0.001$	✓	$<0.001$	✓
Ours (Qwen3.5-27B) vs. Direct VLM (Qwen3.5-27B)	$<0.001$	✓	$<0.001$	✓
Ours (Qwen3.5-4B) vs. Qwen thinking (Qwen3.5-4B)	$<0.001$	✓	$<0.001$	✓
Ours (Qwen3.5-9B) vs. Qwen thinking (Qwen3.5-9B)	$<0.001$	✓	$<0.001$	✓

Appendix

Table 6: Exact paired McNemar tests for the main comparisons (

m=24

tests), with Holm correction within the family. For stochastic conditions, correction is applied to the maximum exact

p

-value across runs.

Comparison	Holm $p$	Sig.	Holm $p$	Sig.
	VirtualHome		ALFWorld
Ours vs. Constrained greedy	1.000	–	$<0.001$	✓
Ours vs. Unconstrained MCTS	1.000	–	$<0.001$	✓

Appendix

Table 7: Exact paired McNemar tests for the component ablations (

m=4

tests). All conditions use Qwen3.5-27B. Holm correction is applied within the family.

Figure 4: Terminal failure attribution without aggregation across environments, Qwen scales, or visual-policy models. Values are percentages of all tasks for the corresponding row. Successful tasks contribute to none of the failure categories.

	VirtualHome			ALFWorld
Condition	4B	9B	27B	4B	9B	27B
Unconstrained greedy	66.0 $\pm$ 0.0	59.0 $\pm$ 0.0	61.0 $\pm$ 0.0	2.4 $\pm$ 0.0	10.0 $\pm$ 0.0	15.4 $\pm$ 0.0
Thinking, temperature zero	47.0 $\pm$ 0.0	31.5 $\pm$ 0.0	59.0 $\pm$ 0.0	5.6 $\pm$ 0.0	10.8 $\pm$ 0.0	16.9 $\pm$ 0.0
Thinking, recommended sampling	39.7 $\pm$ 3.1	31.0 $\pm$ 9.1	57.5 $\pm$ 7.9	5.3 $\pm$ 0.5	14.1 $\pm$ 0.4	18.5 $\pm$ 0.8
Constrained greedy	84.0 $\pm$ 0.0	99.5 $\pm$ 0.0	98.5 $\pm$ 0.0	24.0 $\pm$ 0.0	31.5 $\pm$ 0.0	32.3 $\pm$ 0.0
Unconstrained MCTS	78.0 $\pm$ 0.0	93.0 $\pm$ 0.0	99.0 $\pm$ 0.0	4.8 $\pm$ 0.0	22.3 $\pm$ 0.0	29.2 $\pm$ 0.0

Appendix

Table 8: Task success (%) for every Qwen3.5 model scale and Phase-II condition, reported as accuracy

\pm

standard deviation. Recommended-sampling thinking values use three seeds, and all other conditions use deterministic decoding. Bold denotes the best result in each column and any result not significantly different from it.

Environment	Policy	Model	Grounding	Interaction limit	Model Done
VirtualHome	Direct Qwen	4B	16.0	71.5	7.0
VirtualHome	Direct Qwen	9B	13.0	77.0	0.0
VirtualHome	Direct Qwen	27B	0.5	37.0	0.0
VirtualHome	Embodied	MiMo-7B	10.0	40.0	31.5
VirtualHome	Embodied	R1.5-8B	0.0	50.0	49.0
VirtualHome	Embodied	RoboBrain-32B	13.0	60.0	0.0

Appendix

Table 9: Fine-grained terminal failure mechanisms as percentages of all tasks for each designated main-paper run. Grounding counts terminal mechanisms only. A policy can encounter many failed actions (Table 10 ) without terminating on the ungroundable threshold.

Environment	Policy	Model	Encounters failure	Recovers task	Repeats request	Persistent loop
VirtualHome	Direct Qwen	4B	64.5	5.4	74.5	45.5
VirtualHome	Direct Qwen	9B	53.5	9.3	33.0	79.5
VirtualHome	Direct Qwen	27B	29.0	37.9	8.3	9.5
VirtualHome	Embodied	MiMo-7B	95.0	17.4	59.6	43.5
VirtualHome	Embodied	R1.5-8B	72.5	0.7	4.6	58.0
VirtualHome	Embodied	RoboBrain-32B	93.0	22.0	62.1	50.0

Appendix

Table 10: Failure response of direct and embodied policies (%). Encounters failure is the fraction of tasks containing an ungroundable or failed execution. Recovers task is success among those affected tasks. Repeats request is the fraction of failed-action transitions followed by the same class-level request. Persistent loop is the fraction of all tasks containing at least five identical consecutive requests or a six-step two-action oscillation.

Panel of Figure 5 — Figure 5: Example failure sequences. Each row follows an episode from its initial or first informative state through progress to the decisive error and its terminal consequence. The red callout in (a) identifies the relevant object.

Panel of Figure 6 — Figure 6: Example VirtualHome Phase-I exploration prompt.

Panel of Figure 7 — Figure 7: Example VirtualHome Phase-II planning prompt.

Panel of Figure 8 — Figure 8: VirtualHome direct-interaction prompt components and environment feedback. The current first-person image accompanies each request.

Neuro-Symbolic Planning

See all 20 papers

Plan Right, Then Plan Tight: Symbolic RL for Efficient Embodied Reasoning
Jun 30, 2026Xiangli Shi, Xiaomeng Zhu, Ye Tian +5Large Language Model-Based Robot Planning Symbolic Planning
1Tsinghua University · 2The Hong Kong University of Science and Technology · 3Tencent +1
RoboAgent: Chaining Basic Capabilities for Embodied Task Planning
Apr 9, 2026Peiran Xu, Jiaqi Zheng, Yadong MuRobot Task Planning Embodied Reasoning
Peking University, Beijing, China · XYZ Embodied AI, Beijing, China
Mimir: A Neuro-Symbolic Memory System with Dynamic Grounding for Embodied Agents in Interactive Environments
Aug 5, 2026Haoming Xu, Zhenlin He, Hengyi Wang +2Neuro-Symbolic Reasoning Long-Horizon Agent Tasks
PrimeBot Research Institute · Peking University

Jun 30, 2026J/K move · Enter open · S save

Plan Right, Then Plan Tight: Symbolic RL for Efficient Embodied Reasoning

Xiangli Shi, Xiaomeng Zhu, Ye Tian, Yuchun Guo+4

1Tsinghua University · 2The Hong Kong University of Science and Technology · 3Tencent+1

Large Language Model-Based Robot Planning Symbolic Planning Robot Task Planning Reinforcement Learning with Verifiable Rewards Embodied Reasoning

Open paper AskPDF arXiv

Embodied task planning asks an agent to turn a natural-language instruction into an executable sequence of actions in a physical scene, and is a building block for household, assistive, and service robots. Recent prompting-based and reinforcement-learning planners generate fluent action text but lack a cheap deterministic check that the produced plan is valid in the target world, while high-fidelity simulation is too slow to serve as an inner-loop training signal. The general problem is therefore how to obtain verifiable supervision and rewards for embodied planners without relying on string-level matching or full simulation. Here we show that a single BDDL specification, automatically constructed from open-world video evidence or curated tasks, can serve as a shared interface for data construction, plan verification, and reward design. A video-to-BDDL parser, an LLM verifier, and a lightweight symbolic engine together supply dense feedback at millisecond latency. We further introduce GroupAdapt, a difficulty-aware length schedule that uses the in-batch group pass rate as a zero-cost signal so that hard prompts get wider length tolerance and automatically tighten as their pass rate improves. Under the guidance of the proposed verifier and GroupAdapt schedule, the 8B planner attains a Strict-Pass score of 97.3 on BEHAVIOR-1000, yielding a 25.9 percent relative improvement over the Qwen3-8B baseline. This result exceeds the strongest large-model baseline by 3.5 percent, while simultaneously compressing the response length by 79 percent to 207 tokens, demonstrating both effectiveness and efficiency.

Abstract

Neuro-Symbolic Planning

Explore similar work

Plan Right, Then Plan Tight: Symbolic RL for Efficient Embodied Reasoning

RoboAgent: Chaining Basic Capabilities for Embodied Task Planning

Mimir: A Neuro-Symbolic Memory System with Dynamic Grounding for Embodied Agents in Interactive Environments