Learning to Live: A Step Toward Lifelong Embodied Intelligence

Lirong Che; Peidong Wang; Yimeng Su; Jiong Wang; Yang Chen; Wenbo Fu; Zhenyu Huang; Tiancheng Xu

cs.ROOct 30, 2026

Learning to Live: A Step Toward Lifelong Embodied Intelligence

Authors: Lirong Che, Peidong Wang, Yimeng Su, Jiong Wang, Yang Chen, Wenbo Fu, Zhenyu Huang, Tiancheng Xu,

Organizations: Agibot Human Intelligence Project · AgibotTech/LIFE

Abstract

Embodied agents are usually organized around individual tasks, yet the physical world keeps changing between tasks and during execution. We introduce LIFE (Lifelong Intelligence from Experience), a framework in which a robot observes continuously, remembers selectively, and turns experience into new capabilities. RoboBeats, a family of lightweight streaming gates, tracks conditions set by the agent and discovers new events, deciding when an observation warrants a response or a memory write. Selected observations enter Plastic Memory, a flat store of images and structured records that keeps the source evidence, so new fields and indexes can be added when later questions need them. When existing tools fall short, Explore-Distill-Mount (EDM) turns solutions found during problem solving into validated, reusable tools and skills. Under matched protocols, RoboBeats substantially improves condition tracking over its base models and, at the same writing budget, retains more useful evidence for memory question answering than uniform sampling. A case study of limited scale, replaying 24 hours of continuous video recorded by a real robot, illustrates how the integrated system builds memory from a day of experience and uses it to answer everyday questions. LIFE takes a step toward robots that become more familiar with their surroundings and more capable through experience: learning not only to complete tasks, but to live in the world.

Figures & tables

Figure 1: RoboBeats across streaming perception and memory question answering benchmarks. Results cover temporal localization (Charades-STA), condition triggering (StreamingBench), action-boundary responses (SPOT-Bench), proactive response (ESTP-Bench), and retrospective question answering (OVO-Bench and SSS-Bench). Each panel uses its benchmark-specific metric; higher is better. RoboBeats-2B* uses a OneStreamer-4B reader. Details are given in Section 3.6 .

Figure 2: Overview of LIFE. Continuous perception detects changes in the environment, while active exploration extends observation coverage. Together, they accumulate memories with temporal and spatial information. The agent core combines the current context with historical evidence to reason, plan, and invoke tools, and in turn directs attention and exploration. Autonomous learning develops candidate methods to address capability gaps encountered through experience and adds validated methods to the tool and skill layer.

Figure 3: Everyday experience supports proactive and requested assistance. (a) Observations of snack-taking are retained and used to offer help on a later encounter, followed by fetching and delivery after the user agrees. (b) A request for “my bottle” is interpreted by connecting the user’s statement, observed use, and past sightings to identify the bottle and retrieve its last recorded location.

Figure 4: Overview of RoboBeats. Conditions share a causal visual stream and receive independent predictions. Condition tracking (top right) compares predicted activation intervals with annotations. Event discovery and memory question answering (bottom right) connect selected observations to a writer and reader; future questions are not provided to the gate.

Figure 5: Construction of perception training data and SSS-Bench. (A) Perception training data. Public videos and annotations are screened, grouped by source, deduplicated, and temporally aligned to form continuous clips. Condition-tracking labels follow native action intervals. For event discovery, a teacher identifies windows from event onset to sufficient visual evidence, which are merged into positive intervals. Both branches undergo data balancing, format and timestamp checks, and visual inspection. (B) SSS-Bench. Video evidence supports question construction, verification, and shortcut probing. The kitchen example relates historical evidence to a later question.

Figure 6: From streaming perception to event-grounded question answering. (A) Condition tracking. (B) Event discovery. (C) SSS-Bench questions whose answers depend on earlier observations.

Figure 7: Overview of SSS-Bench. The benchmark contains 618 questions over 171 parent videos from six sources. The panels show example frames, the scenes and actions covered, the source composition, and the number of questions in each of the three question families and 14 subtypes.

Figure 8: Three uses of condition tracking in robot interaction. (a) Timely response: the robot monitors an elevator door and enters when it opens; the API timing shown is from a replay of this example. (b) Task composition: visual conditions connect reach and grip during a bag handover, and trigger pauses and resumption during a phone call. (c) Execution feedback: the grasp condition is monitored while the robot grasps a snack.

Method	Size	EPM $\uparrow$	ASI $\uparrow$	HLD $\uparrow$	BT Avg. $\uparrow$	EPM+ASI Avg. $\uparrow$
Proprietary MLLMs
Gemini 1.5 Pro ( Gemini Team Google, 2024 )	—	58.6	76.4	52.6	62.5	67.5
GPT-4o ( OpenAI, 2024 )	—	57.9	75.7	48.7	60.8	66.8
Open-Source Offline MLLMs
Qwen2-VL-72B ( Wang et al., 2024a )	72B	52.5	60.8	57.5	57.0	56.7
LLaVA-Video-7B ( Zhang et al., 2024 )	7B	56.2	57.4	7.5	40.4	56.8

Table 1: Backward Tracing results on OVO-Bench (%). EPM: episodic memory; ASI: action sequence identification; HLD: hallucination detection. BT Avg. is the equally weighted mean of EPM, ASI, and HLD, except where noted; EPM+ASI Avg. averages EPM and ASI equally. Methods are grouped by how the systems operate. These groups are used for presentation and do not imply equal model sizes, training, or accessible evidence across methods.

Method	Size	Charades-STA [email protected]	StreamingBench PO $\pm 2$ s	SPOT-Bench T-F1@5
Streaming methods, published results, original protocols
StreamingVLM ( Chatterjee et al., 2026 )	7B	–	–	3.9
LiveStar ( Yang et al., 2026b )	8B	–	13.6	–
Dispider ( Qian et al., 2025 )	7B+1.5B	–	25.3	–
MMDuet2 ( Wang et al., 2025d )	3B	–	–	14.2
JoyAI-VL-Interaction	8B+4B	–	–	17.0

Table 2: Results on three condition-tracking benchmarks (%). Metrics are causal online [email protected] (Charades-STA), first-trigger accuracy within

\pm 2

s (StreamingBench PO), and action-boundary T-F1@5 (SPOT-Bench). Published methods use their original protocols; general VLMs and RoboBeats use our condition-judgment protocol. RoboBeats training includes Charades-STA training data. Bold marks the best among our evaluated models with at most 8B parameters; dashes indicate unreported results. Appendix B.3 provides protocol details, full results, and untuned baselines.

Existing Methods
Method	Explicit proactive tasks									Implicit proactive tasks					Context QA			Overall
Method	OR	AP	TRU	OL	OSC	EOL	EOSC	AR	Avg.	OFR	IFR	NAR	TU	Avg.	ORC	TRC	Avg.	Overall
MMDuet	6.7	9.6	17.1	9.8	3.8	5.6	8.6	7.5	8.6	9.5	7.3	48.0	68.5	33.3	17.4 †	23.1 †	20.3 †	17.8 †
Qwen2-VL	12.6	12.0	14.8	28.3	7.1	14.2	15.8	10.1	14.4	16.5	18.7	54.5	61.0	37.7	13.0 †	7.7 †	10.4 †	21.3 †
MiniCPM-V	13.2	15.5	16.2	24.9	6.5	11.7	11.6	12.3	14.0	14.3	20.0	44.2	60.9	34.8	24.6	27.8	26.2	22.9 †
VideoLLM-EyeWO	19.4	24.4	18.7	23.1	16.1	20.4	20.9	19.9	20.4	23.6	26.3	52.4	34.8	34.3	39.4	54.9	47.1	28.2

Table 3: Condition gating and proactive response results on ESTP-Bench. Bold marks the best result within the triggering-strategy comparison. For context question answering and overall results,

\dagger

denotes values cited from the original papers.

Figure 9: Memory question answering by observation duration on SSS-Bench. (a) Accuracy across observation durations. (b) Accuracy differences between RoboBeats-2B and each baseline.

Figure 10: Memory question answering accuracy by question type on SSS-Bench.

Reader	Memory source	EPM $\uparrow$	ASI $\uparrow$	EPM+ASI Avg. $\uparrow$	Gain (pp)
Qwen3.8-27B	OneStreamer PHCM	41.1	59.5	50.3	—
Qwen3.8-27B	RoboBeats-gated	46.8	66.2	56.5	+6.2
OneStreamer-4B	OneStreamer PHCM	44.8	59.5	52.1	—
OneStreamer-4B	RoboBeats-gated	48.1	68.2	58.2	+6.1

Table 4: Text-memory QA across readers (%). Readers use only the question and memory under a fixed answering setup. RoboBeats-gated pairs a RoboBeats-2B gate with a Qwen3.8-27B writer. EPM+ASI Avg. is the two-task mean; Gain is relative to OneStreamer PHCM.

Reader	Memory source	Writer size	Writer input [-1pt](M tokens/h) $\downarrow$	EPM	ASI	EPM+ASI Avg. $\uparrow$
Qwen3.8-27B	All frames	27B	0.338	67.7	72.3	70.0
	Most recent 16 frames	—	0	65.0	63.5	64.2
	OneStreamer PHCM	4B	23.8	65.3	68.2	66.8
	RoboBeats-gated	27B	0.186	62.3	70.9	66.6
OneStreamer-4B	All frames	27B	0.338	65.3	66.9	66.1
	Most recent 16 frames	—	0	62.3	64.2	63.2

Table 5: Writing cost and memory QA. Readers use text memory and 16 recent frames; the recent-frames baseline uses no memory. Writers: 27B = Qwen3.8-27B; 4B = OneStreamer-4B. QA scores are percentages; EPM+ASI Avg. is the two-task mean. Bold compares memory-construction methods within each reader.

Figure 11: Plastic Memory and capability expansion in LIFE. Retained keyframes, initial fields, and extended fields form shared spatiotemporal evidence. The execution agent retrieves evidence through tools; the coding agent develops processing methods in a programmable workspace. EDM organizes exploration outputs into tools, skills, and memory builders for subsequent reuse. This diagram focuses on memory and capability expansion; the complete LIFE architecture is presented in Section 2.

Literature results reported in ReSight
Method	Image $\uparrow$	Object $\uparrow$	Language $\uparrow$	Overall $\uparrow$
DynaMem ( Liu et al., 2025 )	16.1	7.5	5.1	9.6
VLMaps ( Huang et al., 2023 )	43.3	69.7	61.0	58.8
LagMemo ( Zhou et al., 2025 )	56.4	88.4	66.8	70.8
OpenFunGraph ( Zhang et al., 2025a )	53.4	67.1	69.0	63.1
HOV-SG ( Werby et al., 2024 )	–	70.8	82.3	76.5

Table 6: GOAT-core retrieval without EDM: SR@5 (%). Results reproduced from ReSight ( Che et al., 2026 ) . The literature rows were reported in MoM Table I ( Zhou et al., 2026 ) and reproduced in ReSight; HOV-SG’s overall score averages only its two supported query types. Qwen and Sol denote the memory-annotation models; Sol is GPT-5.6-sol. Retrieval-only configurations use no query-time VLM reasoning. Groups differ in provenance and configuration and are not all reruns under a single matched protocol.

Figure 12: Explore–Distill–Mount for a booth photo query in ReSight. Explore inspects historical images and develops photo retrieval. Distill refines candidate tools using answers and traces from the execution agent, consolidating repeated sightings of the same place. Mount registers the tool and its interface. In the illustrated run, one call to the mounted tool returns two distinct booths and supporting frames. This call count is specific to the example; aggregate results appear in Table 9 .

Literature results reported in ReSight
Method	Binary (%) $\uparrow$	Text (%) $\uparrow$	Position error (m) $\downarrow$	$\leq$ 10 m (%) $\uparrow$	Time error (min) $\downarrow$
ReMEmbR-2B ( Anwar et al., 2024 )	43.2	–	53.47	–	2.287
ReMEmbR-8B	46.3	–	55.89	–	4.106
ConceptGraphs ( Gu et al., 2023 )	29.9	–	111.29	–	–
DAAAM-Mini ( Gorlo et al., 2026a )	71.1	–	41.75	–	1.792
Sol execution configurations

Table 7: OC-NaVQA results on 210 queries , reproduced from ReSight ( Che et al., 2026 ) . Mini denotes GPT-5-mini, 2B/8B denote NVILA-Lite models, and Sol denotes GPT-5.6-sol.

\dagger

marks configurations returning the observation-camera pose. The literature rows were reported by DAAAM ( Gorlo et al., 2026a ) and reproduced in ReSight. CodingAgent-Opus-5 uses a separate exploration budget and is shown as a reference, not a matched-budget online comparison. Bold marks the best result among the Sol execution configurations.

Method	Spatial $\uparrow$	Temporal $\uparrow$	Multi-goal $\uparrow$	Overall $\uparrow$	Tokens/q (k) $\downarrow$
Capability-construction episodes: 91–100, 596 queries
Literature results from RAVEN, as reported in ReSight
Random	13.4	6.9	0.5	9.2	–
Embedder Only	31.0	8.1	3.4	18.8	–
VLM Only-Gemini	36.0	28.1	2.2	28.2	–
ReMEmbR-Gemini ( Anwar et al., 2024 )	36.7	17.1	4.5	25.0	–

Table 8: FindingDory capability construction, frozen reuse, and ablation , combining two original ReSight tables ( Che et al., 2026 ) . Success is official high-level success (%); Tokens/q is mean input-plus-output tokens per query in thousands. Gemini denotes Gemini-2.5-Flash and Sol denotes GPT-5.6-sol. The two groups use different episodes and are scored separately. EDM w/o Explore distills only from execution traces, excluding exploration programs, notes, and answers. Online tokens exclude exploration, distillation, and field-construction costs. Bold marks the best result among Sol execution configurations on the capability-construction episodes and among configurations on the held-out episodes.

A. Capability-construction tasks
	Success $\uparrow$		Tool calls $\downarrow$		Tokens/run (k) $\downarrow$
Task	w/o EDM	w/ EDM	w/o EDM	w/ EDM	w/o EDM	w/ EDM
Warmth Corner (8F)	2/5	5/5	8.4	1.0	60.0	10.4
Farthest point (8F)	1/5	5/5	7.2	1.0	38.8	10.0
Booth count (8F)	0/5	5/5	7.6	1.0	50.2	12.9
B. Subsequent tasks: a new query (805) and new memory (1F)

Table 9: Robot-memory query results , reproduced from ReSight ( Che et al., 2026 ) . Each task is run independently five times. Tool calls and input-plus-output tokens are per-run means; the last row reports total successes and mean cost across 30 runs. Grouping distinguishes capability-construction tasks from subsequent tasks.

Figure 13: One day of observations and memory-grounded answers on A3 Ultra. The six cases show cup attribution using additional visual evidence, a cushion handover linked across events, a jacket located using memory and a historical observation pose, an early-morning visitor, the first identified arrival at the workstations, and an estimate of the end of the lunch break refined with historical frames. The agent uses timestamped records and, where needed, revisits visual evidence to answer questions. The jacket question is asked as of 12:30, before its owner came back for it.

Question-level scores
Question / metric	Codex A	Codex B	LIFE
Two cups	$4/6$	$5/6$	$\mathbf{5.5/6}$
Cushion handover and entry	$3/6$	$5.5/6$	$\mathbf{6/6}$
Jacket location (asked at 12:30)	$\mathbf{6/6}$	$\mathbf{6/6}$	$\mathbf{6/6}$
Early-morning visitor	$\mathbf{3/3}$	$\mathbf{3/3}$	$\mathbf{3/3}$
First identified arrival	$2/3$	$\mathbf{3/3}$	$\mathbf{3/3}$

Table 10: Question answering on six real-robot cases.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

General VLMs
Method	Size	Per-source AUROC $\uparrow$								Overall $\uparrow$
Method	Size	Charades	TSU	Edinburgh	IKEA	Perception Test	Ego4D	EGTEA	EPIC	AUROC	AP	Recall
Qwen3-VL-4B	4B	0.726	0.659	0.847	0.447	0.706	0.720	0.784	0.705	0.683	0.738	37.1%
Qwen3-VL-8B	8B	0.727	0.686	0.849	0.404	0.709	0.711	0.811	0.702	0.680	0.739	39.3%
InternVL3.5-8B	8B	0.728	0.668	0.799	0.443	0.724	0.731	0.812	0.728	0.670	0.730	36.2%
MiniCPM-V-4.5	8B	0.727	0.686	0.876	0.407	0.697	0.712	0.780	0.690	0.683	0.732	34.0%

Appendix

Table 11: Overall and per-source results on the held-out condition judgment set.

Offline methods, published results
Method	Readout	[email protected]	[email protected]	[email protected]	mIoU
2D-TAN ( Zhang et al., 2019 )	Offline	57.3	45.8	27.9	41.0
MMN ( Wang et al., 2021 )	Offline	65.4	53.3	31.5	46.5
TimeSuite ( Zeng et al., 2024 )	Offline	69.9	48.7	24.0	–
FV-Action (Qwen2.5-VL-7B) ( Huang et al., 2026 )	Offline	70.1	56.8	29.7	–
UniVTG ( Lin et al., 2023 )	Offline	72.6	60.2	38.6	52.2

Appendix

Table 12: Charades-STA temporal localization results (%). Offline methods are published results that use the complete video. All online rows use our condition-judgment protocol at 1 fps: from causal observations, the model judges at each second whether the query holds, and the same fixed readout converts these judgments into an interval. General VLMs see the most recent eight seconds. RoboBeats training includes Charades-STA training data. Bold marks the best result among our online evaluations.

Published, original protocols: closed-source
Method	$\pm 2$ s	$\pm 1$ s	$\pm 4$ s
Claude Fable 5	71.6	–	–
GPT-6	73.2	–	–
Gemini 3.7 Flash	80.0	–	–
Published, original protocols: open-source
Qwen3.5-4B (ProAgent base)	11.6	–	–

Appendix

Table 13: StreamingBench proactive output (%). Proactive output reports first-trigger accuracy within the indicated time window. Published results use their original protocols. General VLMs, base models, and RoboBeats use our condition-judgment protocol at 1 fps: the model judges at each second whether the condition holds, and the first trigger after the instruction is the response. General VLMs see an eight-second causal window. Bold marks the best result among our evaluations.

Method	Size	T-score@5	T-F1@5	Ego-Exo4D T-F1@5	HTStep T-F1@5
Published Results
StreamingVLM	7B	4.8	3.9	—	—
Qwen2.5-VL	7B	41.8	10.0	—	—
Qwen3-VL	8B	40.5	10.7	—	—
MMDuet2	3B	13.9	14.2	—	—
MiniCPM-o 4.5	9B	—	15.4	—	—

Appendix

Table 14: SPOT-Bench action boundary detection results. All results use the official SPOT harness and scorer with strictly causal input. For general VLMs, base models, and RoboBeats, each request is rewritten as start and end conditions, the model judges them at 2 fps (the input rate of the benchmark’s streaming baselines), and a fixed readout emits the responses. Bold marks the best result among the models we evaluate.

Method	Reader	Overall (%) $\uparrow$	Question type			Observation duration (min)
			Temporal localization (247)	Event order & repetition (202)	Object/person changes (169)	$<10$ (33)	10–15 (154)	15–20 (240)	$\geq 20$ (191)
Native end-to-end answering
JoyAI-VL-Interaction	Native	24.1	7.3	27.7	44.4	42.4	24.0	22.5	23.0
OneStreamer	Native	21.8	9.3	19.3	43.2	36.4	19.5	21.7	21.5
Direct frames, no memory
Recent context (32 s)	Qwen3.8-27B	13.6	1.2	16.3	28.4	39.4	9.1	12.9	13.6

Appendix

Table 15: SSS-Bench question-answering results (%). Native methods answer with their own systems; all other rows use Qwen3.8-27B as the reader. Direct-frame baselines see the sampled frames with no memory and no review. In the last group, the reader receives the text memory and may request one review window of the observation history, from which at most 10 frames are returned. Bold marks the best result within each group.

Text memory only
Reader	Memory source	EPM	ASI	HLD	BT Avg. $\uparrow$
	All frames	51.18	64.19	98.39	71.25
	OneStreamer PHCM	44.78	59.46	99.46	67.90
OneStreamer-4B	RoboBeats-gated	48.15	68.24	98.39	71.59
	All frames	53.20	66.89	88.71	69.60
	OneStreamer PHCM	41.08	59.46	85.48	62.01

Appendix

Table 16: Memory QA comparisons on OVO-Bench. All rows are locally evaluated module configurations. All-frame and RoboBeats-gated memories are written by Qwen3.8-27B; RoboBeats selects the frames in the gated setting. OneStreamer PHCM uses OneStreamer-4B to write memory. All scores are percentages. BT Avg. averages EPM, ASI, and HLD equally. Bold marks the best result among memory sources for each reader and answer-time setting.

Method	BT Avg. $\uparrow$	EPM $\uparrow$	ASI $\uparrow$	HLD $\uparrow$	Input tokens (M) $\downarrow$	Output tokens (M) $\downarrow$
DeltaStreamer	59.19	52.86	68.24	56.45	2,925.2	63.2
DeltaStreamer + RoboBeats	56.00	54.88	64.19	48.92	1,278.0	32.1

Appendix

Table 17: QA performance (%) and token usage of DeltaStreamer. BT Avg. averages EPM, ASI, and HLD equally. Qwen3.8-27B is used for memory writing and answering.

Field	Content
ref	Frame reference identifying the stored image.
t	Capture time.
pose	Capture position, body heading, head rotation, and map. When the writer does not supply a pose, the pose timeline is queried at the capture time.
source , trigger	How the frame was written (route sampling, on-site user confirmation, or proactive perception) and the reason, such as a state change, an event, state maintenance, a fired condition, or a person’s arrival.
event	Activity associated with the write: type, identifier, and text, such as a state or event description, the original condition, or a navigation task.
segment , tracks	Perceptual state segment covering the capture time and the tracks of people present.

Appendix

Table 18: Fields in a keyframe record. All keyframes share these fields. Source and trigger describe how a frame was written; derived fields come from builders applied to every frame.

Continual Robot Learning

Continual robot learning develops robots that acquire new skills and adapt to changing conditions while retaining previously learned behavior.

See all 96 papers

Ask to Evolve: A Unified Harness for Lifelong Embodied Navigation with Hybrid Action Space and Persistent Memory
Oct 29, 2026Kai Sheng, Liuyi Wang, Jinlong Li +3Embodied Navigation Lifelong Learning Agents
School of Electronic and Information Engineering, Tongji University, Shanghai, China
LifeMem: Enabling Lifelong Experience Reuse for LLM Agents
Sep 11, 2026Yuli Qiu, Yutong Li, Wei Su +5Continual Learning for LLM Agents LLM Agent Skill Learning
School of Computer Science and Technology, Beijing Institute of Technology · School of Computer Science and Engineering, Beihang University · Research Center for Social Computing and Interactive Robotics, Harbin Institute of Technology +1
When the Past Shapes Action: Understanding Embodied Memory
Oct 30, 2026Siqi Li, Ziyi Bai, Xinrun Xu +2Agent Memory Long-Term Memory for Embodied Agents
Institute of Software, Chinese Academy of Sciences · University of Chinese Academy of Sciences · Beijing Academy of Artificial Intelligence (BAAI)

Oct 29, 2026J/K move · Enter open · S save

Ask to Evolve: A Unified Harness for Lifelong Embodied Navigation with Hybrid Action Space and Persistent Memory

Kai Sheng, Liuyi Wang, Jinlong Li, Haojie Dai+2

School of Electronic and Information Engineering, Tongji University, Shanghai, China

Embodied Navigation Lifelong Learning Agents Vision-Language Navigation Long-Term Memory for Embodied Agents

Open paper AskPDF arXiv

Giving embodied agents the freedom to act autonomously, initiate interactions, and understand their surroundings through experience is a step toward bringing them to life in the physical world. Realizing this vision requires action, interaction, and memory to work together, so that each encounter can inform current decisions and contribute experience for future tasks. We present EvoLife, a unified navigation harness that enables a multimodal embodied agent to choose how to move, when to ask, and what experience to carry forward. Its hybrid action space supports waypoint navigation, local discrete adjustments, pixel-based navigation, and object-directed navigation. Active interaction allows the agent to seek clarification or assistance when it encounters ambiguity or difficulty, while persistent spatial memory binds observations and dialogue to places, retaining useful experience to guide subsequent visits. We evaluate eight benchmarks across four task families: vision-language navigation, object navigation, interactive instance navigation, and continuous multi-task navigation. With an average of only 1.87 active interactions per episode, our method achieves a first-visit success rate (SR) of 100% across four 100-episode sets spanning vision-language navigation and object navigation. Complete-sequence SR reaches 98% on 50 LHPR-VLN sequences and 94% on 50 GOAT-Bench sequences. In the R2R revisit setting, success weighted by path length (SPL) increases from 82.23% to 93.31% with retained experience. Multi-day physical navigation further demonstrates exploration, proactive help-seeking, and revisiting remembered places. Together, these results advance the vision of bringing embodied agents to life through autonomous action, proactive interaction, and the accumulation and reuse of experience.

Abstract

Continual Robot Learning

Explore similar work

Ask to Evolve: A Unified Harness for Lifelong Embodied Navigation with Hybrid Action Space and Persistent Memory

LifeMem: Enabling Lifelong Experience Reuse for LLM Agents

When the Past Shapes Action: Understanding Embodied Memory