Jaxolotl: A Unified High-Performance Benchmark Suite for LTL-Based Multi-Task RL

Mathias Jackermeier; Jacques Cloete; Alessandro Abate

cs.LGSep 29, 2026

Jaxolotl: A Unified High-Performance Benchmark Suite for LTL-Based Multi-Task RL

Authors: Mathias Jackermeier, Jacques Cloete, Alessandro Abate

Organizations: Department of Computer Science, University of Oxford · Oxford Robotics Institute, University of Oxford

Abstract

Training agents to follow arbitrary instructions is an important goal of multi-task reinforcement learning (RL). Linear temporal logic (LTL) provides a precise and structured formalism for specifying instructions to agents, and has been successfully adopted for training generalist multi-task policies. However, differences in implementations, task distributions, and evaluation protocols make existing methods difficult to compare, while high computational costs limit the scale and statistical reliability of experiments. We introduce Jaxolotl, a unified high-performance benchmark suite for multi-task LTL-RL to address these concerns. Jaxolotl provides a modular, end-to-end JAX implementation of six representative algorithms and four environments, together with newly curated task suites and a standardised, statistically robust evaluation protocol. By precompiling symbolic task representations into static arrays, Jaxolotl enables fully JIT-compiled training and evaluation, achieving end-to-end speedups of up to $220\times$ and supporting controlled comparisons at substantially greater experimental scale. We use this framework to systematically evaluate existing approaches, revealing complementary strengths and limitations: general methods capable of non-myopic reasoning struggle as the number of propositions grows, while methods with stronger scaling rely on environment-specific assumptions and suffer from myopia.

Figures & tables

Figure 1: Overview of Jaxolotl ’s design and implementation. We provide a modular implementation of representative multi-task LTL-RL algorithms in terms of a generic task representation, policy architecture, and shared common modules. Precompilation of dynamic structures and environment resets enables end-to-end JIT compilation of the entire training and evaluation loop.

Panel of Figure 2 — Figure 2: Visualisations of the main environments included in Jaxolotl .

Figure 3: End-to-end training and evaluation times of Jaxolotl compared to original reference implementations on ZoneEnv. We achieve speedups of up to

220\times

for training and

79\times

for evaluation.

Figure 4: Validation of Jaxolotl . (Left) Final ZoneEnv success rates are within

\pm 1.1\%

of the reference implementations for all methods. (Right) Overall training dynamics of DeepLTL and GenZ-LTL closely follow the references.

Figure 5: Aggregate finite-horizon success rates over training. Lines show means across 10 independently trained policies, with shaded

95\%

Student-

t

confidence intervals.

	LetterWorld		ZoneEnv			FrankaZoneEnv		Warehouse
Method	Fin. $\uparrow$	Inf. $\uparrow$	Fin. $\uparrow$	Inf. $\uparrow$	R-S $\uparrow$	Fin. $\uparrow$	Inf. $\uparrow$	Fin. $\uparrow$	Inf. $\uparrow$	R-S $\uparrow$
LTL2Action	0.42 ±0.03	—	0.43 ±0.03	—	—	0.04 ±0.01	—	0.34 ±0.09	—	—
GCRL-LTL	0.94 ±0.00	8.0 ±0.1	0.93 ±0.00	5.7 ±0.1	30.2 ±3.3	0.55 ±0.27	8.1 ±5.9	—	—	—
DeepLTL	0.87 ±0.01	5.3 ±0.2	0.92 ±0.02	3.0 ±0.3	540.1 ±48.0	0.51 ±0.17	4.8 ±3.3	0.80 ±0.09	2.6 ±0.3	618.1 ±118.7
GenZ-LTL †	0.99 ±0.00	8.2 ±0.0	1.00 ±0.00	6.5 ±0.2	492.0 ±50.6	1.00 ±0.00	35.0 ±0.6	—	—	—
$\;\hookrightarrow$ unreduced obs.	0.93 ±0.00	5.6 ±0.0	0.96 ±0.01	5.0 ±0.1	321.6 ±59.2	0.45 ±0.22	5.1 ±3.6	0.84 ±0.12	2.2 ±0.7	102.6 ±24.8

Table 1: Aggregated benchmark results over the Jaxolotl task suites, averaged over 512 episodes per formula, with

95\%

CIs over 10 training seeds. Fin.: finite-horizon success rate (0–1); Inf.: completed accepting cycles with fixed episode length. R-S: completed accepting cycles on the reach-stay task suite. Best results are bold , second best underlined , — denotes unsupported configurations. See Section H.1 for per-specification results.

Figure 6: (Left) Non-myopic reasoning results on different step variants of ConveyorWorldSimple-

k

. (Middle) Aggregate finite-horizon success rates on ZoneEnv-NM over training. (Right) Results of scaling the number of atomic propositions in FrankaZoneEnv.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 7: Environment throughput of Jaxolotl and the reference implementations against the number of parallel environments (log scales). The references run at most 128 parallel environments; the dashed line marks their best measured throughput. Annotations give the peak throughput of Jaxolotl and its speedup over this best reference measurement.

Panel of Figure 8 — Figure 8: (a) A visualisation of ConveyorWorld- $k$ ; the green circle marks the agent start position, and the upper and lower one-way conveyor belts lead to rooms $A$ and $B$ respectively. (b) Non-myopic reasoning results on different step variants of ConveyorWorld- $k$ .

	ID	LTL Formula
LetterWorld	$\varphi_{1}$	$\mathsf{F}\,(\mathsf{a}\wedge(\neg\mathsf{b}\;\mathsf{U}\;\mathsf{c}))\wedge\mathsf{F}\,\mathsf{d}$
	$\varphi_{2}$	$(\mathsf{F}\,\mathsf{d})\wedge(\neg\mathsf{f}\;\mathsf{U}\;(\mathsf{d}\wedge\mathsf{F}\,\mathsf{b}))$
	$\varphi_{3}$	$(\mathsf{F}\,((\mathsf{a}\vee\mathsf{c}\vee\mathsf{j})\wedge\mathsf{F}\,\mathsf{b}))\wedge(\mathsf{F}\,(\mathsf{c}\wedge\mathsf{F}\,\mathsf{d}))\wedge\mathsf{F}\,\mathsf{k}$
	$\varphi_{4}$	$\neg\mathsf{a}\;\mathsf{U}\;(\mathsf{b}\wedge(\neg\mathsf{c}\;\mathsf{U}\;(\mathsf{d}\wedge(\neg\mathsf{e}\;\mathsf{U}\;\mathsf{f}))))$
	$\varphi_{5}$	$\mathsf{G}\,\neg(\mathsf{k}\vee\mathsf{l})\wedge\mathsf{F}\,(\mathsf{a}\wedge\mathsf{F}\,(\mathsf{e}\wedge\mathsf{F}\,(\mathsf{i}\wedge\mathsf{F}\,\mathsf{d})))$
	$\varphi_{6}$	$\mathsf{G}\,\neg(\mathsf{a}\vee\mathsf{b}\vee\mathsf{c}\vee\mathsf{d})\wedge\mathsf{G}\,(\mathsf{h}\rightarrow\neg\mathsf{f}\;\mathsf{U}\;\mathsf{g})\wedge\mathsf{F}\,\mathsf{g}\wedge\mathsf{F}\,(\mathsf{h}\wedge\mathsf{F}\,\mathsf{i})$

Appendix

Table 2: Finite-horizon LTL specifications used for evaluation.

	ID	LTL Formula
LetterWorld	$\psi_{1}$	$\mathsf{G}\,\mathsf{F}\,\mathsf{a}\wedge\mathsf{G}\,\mathsf{F}\,\mathsf{b}\wedge\mathsf{G}\,\mathsf{F}\,\mathsf{c}$
	$\psi_{2}$	$\mathsf{G}\,\mathsf{F}\,(\mathsf{a}\wedge\mathsf{F}\,\mathsf{d})\wedge\mathsf{G}\,\mathsf{F}\,(\mathsf{b}\wedge\mathsf{F}\,\mathsf{e})$
	$\psi_{3}$	$\mathsf{G}\,\neg\mathsf{l}\wedge\mathsf{G}\,\mathsf{F}\,\mathsf{c}\wedge\mathsf{G}\,\mathsf{F}\,\mathsf{g}\wedge\mathsf{G}\,\mathsf{F}\,\mathsf{j}$
	$\psi_{4}$	$(\neg\mathsf{a}\;\mathsf{U}\;\mathsf{b})\wedge\mathsf{G}\,\mathsf{F}\,\mathsf{a}\wedge\mathsf{G}\,\mathsf{F}\,\mathsf{f}\wedge\mathsf{G}\,\mathsf{F}\,\mathsf{g}\wedge\mathsf{G}\,\mathsf{F}\,\mathsf{h}$
	$\psi_{5}$	$\mathsf{G}\,(\mathsf{a}\rightarrow\mathsf{F}\,\mathsf{d})\wedge\mathsf{G}\,\mathsf{F}\,\mathsf{a}\wedge\mathsf{G}\,\mathsf{F}\,\mathsf{h}$
	$\psi_{6}$	$\mathsf{G}\,\mathsf{F}\,(\mathsf{c}\wedge(\neg\mathsf{d}\;\mathsf{U}\;\mathsf{e}))\wedge\mathsf{G}\,\mathsf{F}\,\mathsf{b}$

Appendix

Table 3: Infinite-horizon LTL specifications used for evaluation.

	ID	LTL Formula
ZoneEnv	$\chi_{1}$	$\mathsf{F}\,\mathsf{G}\,\mathsf{red}$
	$\chi_{2}$	$\mathsf{F}\,\mathsf{G}\,\mathsf{red}\wedge\mathsf{F}\,(\mathsf{yellow}\wedge\mathsf{F}\,\mathsf{green})$
	$\chi_{3}$	$\mathsf{F}\,\mathsf{G}\,\mathsf{purple}\wedge\mathsf{G}\,\neg\mathsf{yellow}$
	$\chi_{4}$	$\mathsf{G}\,((\mathsf{green}\vee\mathsf{yellow})\rightarrow\mathsf{F}\,\mathsf{red})\wedge\mathsf{F}\,\mathsf{G}\,(\mathsf{green}\vee\mathsf{purple})$
	$\chi_{5}$	$\mathsf{F}\,\mathsf{G}\,\mathsf{green}$
	$\chi_{6}$	$\mathsf{F}\,\mathsf{G}\,\mathsf{yellow}$

Appendix

Table 4: Reach-stay LTL specifications used for evaluation.

	$\varphi$	LTL2Action	GCRL-LTL	DeepLTL	GenZ-LTL †	GenZ-LTL (unreduced obs.)	SemLTL	StructLTL
LetterWorld	$\varphi_{1}$	0.62 ±0.14	0.99 ±0.00	1.00 ±0.00	0.99 ±0.00	1.00 ±0.00	1.00 ±0.00	1.00 ±0.00
	$\varphi_{2}$	0.74 ±0.15	0.99 ±0.00	0.96 ±0.01	1.00 ±0.00	0.98 ±0.00	0.95 ±0.01	0.97 ±0.01
	$\varphi_{3}$	0.45 ±0.13	1.00 ±0.00	1.00 ±0.00	1.00 ±0.00	1.00 ±0.00	1.00 ±0.00	1.00 ±0.00
	$\varphi_{4}$	0.52 ±0.20	0.98 ±0.01	0.95 ±0.01	0.99 ±0.00	0.97 ±0.01	0.94 ±0.01	0.95 ±0.01
	$\varphi_{5}$	0.02 ±0.01	0.92 ±0.01	0.73 ±0.03	1.00 ±0.00	0.85 ±0.02	0.57 ±0.07	0.79 ±0.04
	$\varphi_{6}$	0.00 ±0.00	0.72 ±0.01	0.53 ±0.07	0.97 ±0.01	0.77 ±0.02	0.32 ±0.04	0.62 ±0.03

Appendix

Table 5: Per-specification finite-horizon benchmark results (success rate), averaged over 512 evaluation episodes, with

95\%

Student-

t

CIs over training seeds. Best results are bold , second best underlined , — denotes unsupported configurations.

	$\psi$	GCRL-LTL	DeepLTL	GenZ-LTL †	GenZ-LTL (unreduced obs.)	SemLTL	StructLTL
LetterWorld	$\psi_{1}$	9.4 ±0.1	6.8 ±0.3	9.8 ±0.1	6.9 ±0.1	6.0 ±0.2	6.7 ±0.4
	$\psi_{2}$	6.4 ±0.0	4.3 ±0.2	6.7 ±0.1	4.5 ±0.0	4.1 ±0.1	4.3 ±0.2
	$\psi_{3}$	8.8 ±0.5	5.3 ±0.4	9.8 ±0.1	6.2 ±0.1	4.5 ±0.3	5.7 ±0.3
	$\psi_{4}$	6.3 ±0.0	4.0 ±0.3	6.7 ±0.1	4.5 ±0.0	3.3 ±0.2	4.1 ±0.2
	$\psi_{5}$	9.4 ±0.3	6.3 ±0.9	6.4 ±0.3	5.2 ±0.3	6.0 ±0.3	7.1 ±0.8
	$\psi_{6}$	9.3 ±0.1	6.4 ±0.5	9.9 ±0.1	6.8 ±0.2	4.2 ±0.6	6.8 ±0.4

Appendix

Table 6: Per-specification infinite-horizon benchmark results (completed accepting cycles), averaged over 512 evaluation episodes, with

95\%

Student-

t

CIs over training seeds. Best results are bold , second best underlined , — denotes unsupported configurations.

	$\chi$	GCRL-LTL	DeepLTL	GenZ-LTL †	GenZ-LTL (unreduced obs.)	SemLTL	StructLTL
ZoneEnv	$\chi_{1}$	24.2 ±5.0	645.1 ±85.6	548.0 ±57.2	340.6 ±119.3	615.7 ±38.5	608.4 ±61.8
	$\chi_{2}$	51.2 ±16.2	497.2 ±61.5	471.7 ±61.7	307.4 ±99.8	445.5 ±22.1	501.2 ±56.8
	$\chi_{3}$	21.6 ±3.3	512.8 ±79.5	533.1 ±51.9	317.6 ±103.0	591.4 ±67.9	575.4 ±63.8
	$\chi_{4}$	28.4 ±6.7	520.9 ±85.7	495.6 ±37.5	303.1 ±104.9	285.3 ±101.4	571.0 ±64.9
	$\chi_{5}$	25.9 ±4.3	635.4 ±51.3	541.4 ±51.2	333.8 ±129.0	679.1 ±64.9	593.8 ±100.8
	$\chi_{6}$	18.9 ±2.6	513.4 ±113.3	299.4 ±94.5	341.8 ±104.1	639.0 ±66.7	626.8 ±97.9

Appendix

Table 7: Per-specification reach-stay benchmark results (completed accepting cycles), averaged over 512 evaluation episodes, with

95\%

Student-

t

CIs over training seeds. Best results are bold , second best underlined , — denotes unsupported configurations.

Category	Hyperparameter	LTL2Action	SemLTL	DeepLTL	StructLTL	GCRL-LTL	GenZ-LTL
PPO	Total environment steps	$2\times 10^{7}$
	Environments	16
	Steps/update	128
	Minibatches	8
	Update epochs	8
	Discount ( $\gamma$ )	0.94

Appendix

Table 8: Training and model hyperparameters for LetterWorld. Values spanning multiple columns are shared by those methods; N/A marks settings that do not apply to a method.

Category	Hyperparameter	LTL2Action	SemLTL	DeepLTL	StructLTL	GCRL-LTL	GenZ-LTL
PPO	Total environment steps	$2\times 10^{7}$
	Environments	16
	Steps/update	4,096
	Minibatches	32
	Update epochs	10
	Discount ( $\gamma$ )	0.998

Appendix

Table 9: Training and model hyperparameters for ZoneEnv and ZoneEnv-NM. Values spanning multiple columns are shared by those methods; N/A marks settings that do not apply to a method.

Category	Hyperparameter	LTL2Action	SemLTL	DeepLTL	StructLTL	GCRL-LTL	GenZ-LTL
PPO	Total environment steps	$2\times 10^{8}$
	Environments	2,048
	Steps/update	64
	Minibatches	4
	Update epochs	5
	Discount ( $\gamma$ )	0.995

Appendix

Table 10: Training and model hyperparameters for FrankaZoneEnv. Values spanning multiple columns are shared by those methods; N/A marks settings that do not apply to a method. The table covers FrankaZoneEnv-

k

for

k\in\{8,9,10,11,12\}

. Hyperparameters that vary with

k

are given as functions of

k

Category	Hyperparameter	LTL2Action	SemLTL	DeepLTL	StructLTL	GenZ-LTL
PPO	Total environment steps	$2\times 10^{8}$
	Environments	1,024
	Steps/update	128
	Minibatches	4
	Update epochs	5
	Discount ( $\gamma$ )	0.998

Appendix

Table 11: Training and model hyperparameters for Warehouse. Values spanning multiple columns are shared by those methods; N/A marks settings that do not apply to a method.

Category	Hyperparameter	LTL2Action	SemLTL	DeepLTL	StructLTL	GCRL-LTL	GenZ-LTL
PPO	Total environment steps	$2\times 10^{6}$
	Environments	256
	Steps/update	64
	Minibatches	4
	Update epochs	8
	Discount ( $\gamma$ )	0.9

Appendix

Table 12: Training and model hyperparameters for ConveyorWorld and ConveyorWorldSimple. Values spanning multiple columns are shared by those methods; N/A marks settings that do not apply to a method. The table covers ConveyorWorld-

k

for

k\in\{1,2,3,4,5,6,7,8,9,10\}

and ConveyorWorldSimple-

k

for

k\in\{1,2,4,8,16,32\}

. Hyperparameters that vary with

k

are given as functions of

k

. Note that environment resets are not precomputed for these environments (since they are trivially cheap), and the training “curriculum” only consists of a single stage (i.e. no curriculum) for all methods.

Explore similar work

PlatoLTL: Scaling LTL-Guided Multi-Task RL
Jan 30, 2026Jacques Cloete, Mathias Jackermeier, Ioannis Havoutis +1Multi-Turn Reinforcement Learning
Oxford Robotics Institute University of Oxford Oxford, United Kingdom · Department of Computer Science University of Oxford Oxford, United Kingdom
JaxMARL: Multi-Agent RL Environments and Algorithms in JAX
Nov 16, 2023Alexander Rutherford, Benjamin Ellis, Matteo Gallici +18Multi-Agent Reinforcement Learning Cpu-Gpu Hybrid Designs
University of Oxford · Universitat Politècnica de Catalunya · University College London +2
Live LTL Progress Tracking: Towards Task-Based Exploration
Apr 18, 2026Noel Brindise, Cedric Langbort, Melkior OrnikLinear Temporal Logics Trajectory Generation
Coordinated Science Laboratory, University of Illinois Urbana-Champaign, Urbana, USA

Jan 30, 2026, cs.LGJ/K move · Enter open · S save

PlatoLTL: Scaling LTL-Guided Multi-Task RL

Jacques Cloete, Mathias Jackermeier, Ioannis Havoutis, Alessandro Abate

Oxford Robotics Institute University of Oxford Oxford, United Kingdom · Department of Computer Science University of Oxford Oxford, United Kingdom

Multi-Turn Reinforcement Learning

Open paper AskPDF arXiv

Linear temporal logic (LTL) has emerged as a powerful formalism for specifying structured, temporally extended tasks in multi-task reinforcement learning (RL). However, while existing approaches in LTL-guided multi-task RL demonstrate success in simple environments, they suffer from challenges in representation learning and exploration efficiency when applied to high-dimensional environments with parameterized specifications. We present PlatoLTL, which elevates state-of-the-art methods to address both challenges. We model atomic propositions as instances of atomic predicates and inject task parameters directly into the goal embedding to enable efficient generalization. We also leverage simple priors on closeness to predicate satisfaction to enable and accelerate learning of complex tasks without biasing the optimal policy. We validate our approach on challenging environments including robotic manipulation and multi-drone navigation.

Abstract

Explore similar work

PlatoLTL: Scaling LTL-Guided Multi-Task RL

JaxMARL: Multi-Agent RL Environments and Algorithms in JAX

Live LTL Progress Tracking: Towards Task-Based Exploration