Organizations: Nanyang Technological University · Beihang University · National University of Singapore · Cloud Butterfly Technology · Shanghai Jiao Tong University
A foundation model should not act in isolation as an embodied agent. Yet, existing methods often optimize individual components of the agent stack, such as memory, context, skills, or action interfaces, rather than treating the supporting system itself as a unified policy. Moreover, interaction alone does not yield self-improvement unless execution experience is converted into persistent, validated system changes. We therefore propose RoboFoundry, the first embodied agentic framework that formulates this process as Self-Evolving System-as-Policy. RoboFoundry diagnoses capability gaps in decision-making and memory management, converts execution traces into validated task-specific system updates, and promotes recurring improvements to the general system. Evolution operates over two complementary surfaces: a context system that manages active internal context and persistent file-system memory, and a hierarchical skill system that organizes atomic skills, reusable compositions, and failure-conditioned recovery. A shared semantic interface separates embodiment-invariant decisions from embodiment-specific execution, allowing evolved system capabilities to transfer across heterogeneous robots. On EmbodiedBench, RoboFoundry achieves state-of-the-art performance, notably improving GPT-5.5 by 27.8%. It also brings Qwen3.7-Plus to near parity with GPT-5.5 (70.3% vs. 72.7%), showing consistent gains from system-as-policy evolution across foundation models. For long-horizon memory, RoboFoundry outperforms all baselines on RoboMemArena by at least 39.0%, even against methods assisted by external foundation models. On LIBERO-PRO, it further outperforms Cap-Agent0 by 243.8%-679.7% across all perturbation types. In real-world deployments, RoboFoundry demonstrates zero-shot transfer and online evolution across robots and tasks, highlighting its potential for fully autonomous embodied agents.
Figures & tables
Figure 1: RoboFoundry: System-as-Policy Evolution across benchmarks and real robots. The central Act–Reflect–Repair–Promote loop executes across embodiments, attributes failures to capabilities, revises the task-level system, and promotes validated changes to the general system. The surrounding panels show gains across foundation models on EmbodiedBench, long-horizon memory results on RoboMemArena, and robustness perturbations under LIBERO-PRO. The capability plots compare GPT-5.5 and Qwen3.7-Plus before and after RoboFoundry, while the lower image strip illustrates adaptation and evolution on physical robots.
Figure 2: Overview of RoboFoundry. (a) The inner loop runs task-specific System-as-Policy through context/skill surfaces and semantic/execution bindings, producing embodiment-specific rollouts and traces. (b) The general system supplies reusable context and skill rules to task-level systems. (c) The outer loop diagnoses capability failures and validates targeted edits through held-in repair and held-out retention. Accepted edits update the task-specific system or promote the general system.
Figure 3: Example of evolution trace on the EB-Habitat spatial subset. Here, RoboFoundry (GPT-5.5) rises Held-in success rate from 21.4% to 78.0% over six candidate evaluations. The final version passes the held-out check and is promoted to the general system.
Figure 4: Example of scaling with held-in data on EmbodiedBench. RoboFoundry (GPT-5.5) is evaluated on 50 fixed held-out tasks per suite. Using three quarters of the held-in data already matches the full-data success rate of 74.0% on EB-Habitat, highlighting data-efficient evolution.
Method
Avg.
EB- ALFRED
EB- Habitat
EB- Navigation
EB- Manipulation
RoboFoundry (GPT-6 Astra)
78.0 +6.1
90.0 +2.7
86.7 +11.4
80.2 +3.2
54.9 +6.9
RoboFoundry (GPT-5.5)
72.7 +15.8
84.0 +7.3
88.7 +24.7
72.0 +17.3
45.9 +13.8
RoboFoundry (Qwen3.7-Plus)
70.3 +10.4
81.3 +8.0
80.3 +15.3
71.9 +7.6
47.7 +10.5
RoboFoundry (GLM5.3-Flash)
69.3 +7.4
80.0 +8.3
75.0 +14.3
72.7 +1.7
49.5 +5.1
RoboFoundry (Qwen3.8-27B)
66.5 +14.2
78.0 +12.3
77.0 +19.0
68.3 +16.0
42.8 +9.5
RoboFoundry-Lite (GPT-5.5)
65.5 +8.6
78.0 +1.3
74.3 +10.3
69.7 +15.0
39.9 +7.8
Table 1: Overall comparison on EmbodiedBench. Avg. is the arithmetic mean of success rates across the four suites. All results are reported in percentage (%). Gains are reported in percentage points (+pp).
Method
Overall
Transfer
Occlusion
Counting
Sequence
RoboFoundry
53.5 / 72.8
70.5 / 78.1
38.3 / 57.6
55.2 / 82.5
75.2 / 92.1
PrediMem
38.5 / 55.2
22.5 / 45.2
27.3 / 38.4
45.7 / 69.3
72.5 / 89.5
MemER
27.3 / 49.1
20.0 / 36.1
16.4 / 33.2
27.1 / 65.1
65.0 / 79.1
HiF-VLA
16.9 / 39.8
17.5 / 38.9
12.7 / 27.1
8.6 / 45.9
42.5 / 70.2
π0.5
21.5 / 38.7
20.0 / 42.8
12.7 / 17.2
14.3 / 50.9
60.0 / 71.6
MemoryVLA
15.0 / 35.3
15.0 / 37.2
7.3 / 13.1
14.3 / 55.1
37.5 / 65.2
Table 2: Long-horizon memory results on RoboMemArena. RoboFoundry achieves the best TSR and CSR overall and across all four categories. Each entry reports TSR / CSR (%).
Table 7
Figure 5: Real-world evolution and transfer of RoboFoundry. (a) Failures motivate alignment and progress-tracking edits, enabling nesting-doll completion in Trial 3. (b) Towel-folding skills evolved on the blue towel transfer to unseen green and pink fabrics. (c) The general system grounds a semantic navigation goal in executable actions on a Unitree G1 humanoid.
Figure 6: Long-horizon chemistry manipulation with RoboFoundry. A task-state record connects six dependent subtasks: funnel readiness and scale taring are written during execution, and the final pouring step reads the earlier funnel-ready state to guide skill execution.
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 7: Representative EmbodiedBench cases (GPT-5.5) before and after RoboFoundry evolution. Examples illustrate repairs to perception, reasoning, planning, and low-level skill execution.
Figure 8: RoboFoundry rollouts across four memory categories in RoboMemArena: Transfer, Occlusion, Counting, and Sequence.
Method
Avg.
Base
Common Sense
Complex Instruction
Visual
Spatial
Long Horizon
RoboFoundry (GPT-6 Astra)
90.0
92.0
92.0
92.0
86.0
94.0
84.0
RoboFoundry (GPT-5.5)
84.0
90.0
84.0
84.0
74.0
80.0
92.0
RoboFoundry (Qwen3.7-Plus)
81.3
88.0
90.0
72.0
74.0
82.0
82.0
RoboFoundry (GLM5.3-Flash)
80.0
82.0
76.0
82.0
80.0
74.0
86.0
RoboFoundry (Qwen3.8-27B)
78.0
82.0
84.0
82.0
72.0
68.0
80.0
RoboFoundry-Lite (GPT-5.5)
78.0
82.0
80.0
76.0
70.0
72.0
88.0
Appendix
Table 5: Main comparison on the EB-ALFRED high-level benchmark. Avg. denotes the arithmetic mean over Base, Common Sense, Complex Instruction, Visual, Spatial, and Long Horizon. All results are reported in percentage (%).
Method
Avg.
Base
Common Sense
Complex Instruction
Visual
Spatial
Long Horizon
RoboFoundry (GPT-5.5)
88.7
100.0
78.0
80.0
86.0
100.0
88.0
RoboFoundry (GPT-6 Astra)
86.7
100.0
100.0
100.0
100.0
42.0
78.0
RoboFoundry (Qwen3.7-Plus)
80.3
100.0
80.0
92.0
88.0
48.0
74.0
RoboFoundry (Qwen3.8-27B)
77.0
100.0
58.0
80.0
80.0
94.0
50.0
RoboFoundry (GLM5.3-Flash)
75.0
100.0
58.0
66.0
80.0
94.0
52.0
RoboFoundry-Lite (GPT-5.5)
74.3
100.0
66.0
74.0
80.0
44.0
82.0
Appendix
Table 6: Main comparison on the EB-Habitat high-level benchmark. Avg. denotes the arithmetic mean over Base, Common Sense, Complex Instruction, Visual, Spatial Relationship, and Long Horizon. All results are reported in percentage (%).
Method
Avg.
Base
Common Sense
Complex Instruction
Visual
Long Horizon
RoboFoundry (GPT-6 Astra)
80.3
90.0
85.0
86.7
73.3
66.7
RoboFoundry (GLM5.3-Flash)
72.7
80.0
71.7
78.3
73.3
60.0
RoboFoundry (GPT-5.5)
72.0
80.0
83.3
80.0
76.7
40.0
RoboFoundry (Qwen3.7-Plus)
72.0
78.3
83.3
65.0
66.7
66.7
RoboFoundry-Lite (GPT-5.5)
69.7
83.3
81.7
76.7
80.0
26.7
RoboFoundry (Qwen3.8-27B)
68.3
76.7
85.0
73.3
63.3
43.3
Appendix
Table 7: Main comparison on the EB-Navigation benchmark. Avg. denotes the arithmetic mean over Base, Common Sense, Complex Instruction, Visual, and Long Horizon. All results are reported in percentage (%).
Method
Avg.
Base
Common Sense
Complex Instruction
Visual
Spatial
RoboFoundry (GPT-6 Astra)
54.8
58.3
62.5
56.3
55.6
41.7
RoboFoundry (GLM5.3-Flash)
49.6
56.3
45.8
56.3
47.2
41.7
RoboFoundry (Qwen3.7-Plus)
47.8
43.8
54.2
52.1
44.4
43.8
RoboFoundry (GPT-5.5)
46.1
54.2
47.9
43.8
41.7
41.7
RoboFoundry (Qwen3.8-27B)
42.5
35.4
41.7
43.8
47.2
45.8
RoboFoundry-Lite (GPT-5.5)
39.5
43.8
37.5
31.3
47.2
39.6
Appendix
Table 8: Main comparison on the EB-Manipulation benchmark. Avg. denotes the success rate over all 228 tasks. All results are reported in percentage (%).
Method
Avg.
EB- ALFRED
EB- Habitat
EB- Navigation
EB- Manipulation
RoboFoundry (GPT-5.5)
71.7 +8.3
82.3 +3.3
81.3 +12.3
75.7 +8.4
47.3 +9.0
RoboFoundry (Qwen3.7-Plus)
70.4 +5.8
82.5 +2.8
79.3 +5.6
71.8 +1.0
47.8 +13.7
RoboFoundry (Qwen3.8 27B)
69.4 +5.1
79.7 +3.4
79.3 +8.0
75.0 +6.0
43.7 +3.0
RoboFoundry (GLM5.3-Flash)
68.8 +7.6
80.0 +5.7
72.0 +5.3
77.7 +8.7
45.5 +10.5
Qwen3.7-Plus
64.6
79.7
73.7
70.8
34.1
Qwen3.8-27B
64.3
76.3
71.3
69.0
40.7
Appendix
Table 9: Overall comparison on EmbodiedBench (16K). Avg. is the arithmetic mean of success rates across the four suites. All results are reported in percentage (%). Gains are reported in percentage points (+pp).
Task (Symbolic Form)
OpenVLA
π0
π0.5
CaP-Agent0
RoboFoundry
RoboFoundry †
Pos
Task
Pos
Task
Pos
Task
Pos
Task
Pos
Task
Pos
Task
Place(alphabet_soup,basket)
0.0
0.0
0.0
0.0
0.0
0.0
2.0
4.0
98.0
99.0
100.0
100.0
Place(bbq_sauce,basket)
0.0
0.0
0.0
0.0
100.0
2.0
12.0
42.0
97.0
93.0
100.0
100.0
Place(butter,basket)
0.0
0.0
0.0
0.0
54.0
0.0
26.0
18.0
97.0
99.0
100.0
100.0
Place(chocolate_pudding,basket)
0.0
0.0
0.0
0.0
0.0
2.0
18.0
48.0
88.0
93.0
100.0
100.0
Place(cream_cheese,basket)
0.0
0.0
10.0
0.0
0.0
0.0
12.0
6.0
98.0
99.0
100.0
100.0
Appendix
Table 10: Task-wise LIBERO-PRO performance on the libero-object benchmark. Each entry reports position/task success rate in percentage (%). † denotes the use of privileged simulator object poses. Results for OpenVLA, π0 , π0.5 , and CaP-Agent0 are taken directly from the CaP-Agent0 report [ Fu et al., 2026 ] .
Task (Symbolic Form)
OpenVLA
π0
π0.5
CaP-Agent0
RoboFoundry
RoboFoundry †
Pos
Task
Pos
Task
Pos
Task
Pos
Task
Pos
Task
Pos
Task
Open(cabinet,drawer_mid)
0.0
0.0
0.0
0.0
0.0
4.0
0.0
0.0
71.0
71.0
95.0
100.0
Put(bowl,drawer_top)
0.0
0.0
0.0
0.0
94.0
2.0
4.0
0.0
90.0
71.0
95.0
100.0
Push(plate,stove_front)
0.0
0.0
0.0
0.0
0.0
0.0
0.0
10.0
71.0
70.0
95.0
100.0
Put(bowl,plate)
0.0
0.0
0.0
0.0
0.0
2.0
36.0
38.0
90.0
97.0
100.0
100.0
Put(bowl,stove)
0.0
0.0
0.0
0.0
0.0
4.0
22.0
12.0
99.0
100.0
95.0
100.0
Appendix
Table 11: Task-wise LIBERO-PRO performance on the libero-goal benchmark. Each entry reports position/task success rate in percentage (%). † denotes the use of privileged simulator object poses. Results for OpenVLA, π0 , π0.5 , and CaP-Agent0 are taken directly from the CaP-Agent0 report [ Fu et al., 2026 ] .
Task (Symbolic Form)
OpenVLA
π0
π0.5
CaP-Agent0
RoboFoundry
RoboFoundry †
Pos
Task
Pos
Task
Pos
Task
Pos
Task
Pos
Task
Pos
Task
Pick(between(plate,ramekin),plate)
0.0
0.0
0.0
0.0
2.0
0.0
22.0
14.0
98.0
88.0
100.0
100.0
Pick(table_center,plate)
0.0
0.0
0.0
0.0
0.0
2.0
22.0
14.0
89.0
99.0
100.0
100.0
Pick(drawer_top(cabinet_wood),plate)
0.0
0.0
0.0
0.0
0.0
0.0
2.0
10.0
88.0
96.0
99.0
98.0
Pick(next_to(cookie_box),plate)
0.0
0.0
0.0
0.0
0.0
2.0
0.0
10.0
97.0
96.0
100.0
98.0
Pick(next_to(plate),plate)
0.0
0.0
0.0
0.0
0.0
0.0
10.0
20.0
97.0
88.0
100.0
98.0
Appendix
Table 12: Task-wise LIBERO-PRO performance on the libero-spatial benchmark. Each entry reports position/task success rate in percentage (%). † denotes the use of privileged simulator object poses. Results for OpenVLA, π0 , π0.5 , and CaP-Agent0 are taken directly from the CaP-Agent0 report [ Fu et al., 2026 ] .
Current embodied systems largely rely on pretrained capabilities that remain fixed after deployment, limiting their ability to learn from physical interaction. We introduce MachEmbodied-Brain (ME-Brain), a self-evolving embodied system organized around a closed loop of action execution, experience acquisition, experience evolution, and improved execution. Evolvable Memory consolidates multimodal trajectories into hierarchical, reusable experience; Cognitive Core transforms physical experience into transferable skills; and the Action Model combines event-driven keyframes, EventCell local-world prediction, and action-conditioned memory modulation to focus computation on decision-critical moments, regions, and historical evidence. Together, these modules shift embodied intelligence from train-and-freeze to deploy-and-evolve without model retraining. Cognitive Core outperforms the strongest comparison models by 8.2 and 9.6 points on embodied and agent benchmarks. The Action Model achieves 47.88% mean success on RoboMME, a 3.26-point improvement over the strongest baseline. On RoboDojo, it reaches a 21.51 mean Score and 16.03% success rate, exceeding π0.5 by 10.10 and 9.12 points. On the six-task ME-RealBench, ME-Brain achieves a 69.5 mean Score and 66.7% success rate, outperforming DM0.5 by 12.8 and 11.7 points, respectively.
Robot foundation models provide strong visuomotor control, yet their performance can degrade when object positions or task instructions change. Further improvements often require post-training on substantial robot data, which can be costly to collect through methods such as teleoperation. Agentic harnesses can adapt around the model, but current self-evolving harnesses use robot trials inefficiently when deciding which code and skill changes to pursue. We introduce EmbodiedRSI, a self-evolving agentic harness that autonomously decides where to explore next and turns the resulting physical interaction into improved code and skills. EmbodiedRSI realizes this through a Fast-Slow Dual-System Architecture, in which competing code and skill hypotheses are maintained in a Hypothesis Graph. Value-of-Information Experiment Selection chooses physical experiments that can distinguish these hypotheses. Their outcomes guide Code-Skill Co-Evolution. The Slow System builds Hierarchical Memory, and Reward-Grounded Memory Learning selects effective memory according to their value for later Fast-System improvement. On RoboCasa365, EmbodiedRSI reaches 77.0% overall success and 71.3% on Composite-Unseen, compared with 40.1% for the best baseline. EmbodiedRSI also reaches 86.8% overall success on LIBERO-Pro. Beyond benchmark performance, EmbodiedRSI transfers zero-shot to real-world robot, achieving 71.3% overall success across multiple challenging tasks.
Embodied agents are increasingly built as systems around foundation models, where performance depends not only on model weights but also on the skills, context, action interfaces, and execution harness surrounding the model. While supervised fine-tuning and reinforcement learning can adapt agents to new environments, they require additional data, rewards, and training runs; meanwhile, many train-free code-centric approaches rely on programmable robot APIs that may be unavailable in fixed-interface settings. We propose SHAPER, a self-evolving framework for train-free embodied adaptation that keeps model parameters frozen and improves the non-parametric agent system by evolving reusable skills and a context-code harness through target-environment rollouts. In SHAPER, the same frozen model can serve as both planner and optimizer, refining its external skills and context-code harness without parameter updates. We evaluate SHAPER on VLABench and ESI-Bench, covering embodied agents with different low-level action interfaces, and compare against pure execution, supervised fine-tuning, and test-time-scaling baselines such as verifier-free selection and voting. Our results suggest that skill-and-harness optimization is a practical route to self-evolving embodied agents when model training is expensive, unavailable, or undesirable.