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.
Figures & tables
Figure 1: The EmbodiedRSI Self-Evolution Loop. Physical outcomes guide code and skill updates while the robot foundation model remains frozen.
Figure 2: Overview of EmbodiedRSI. The Fast System uses a Hypothesis Graph to choose each physical experiment. Each result guides a joint code-skill update. The Slow System retains memories that improve later adaptation.
Figure 3: Hypothesis Graph Updates from Physical Trials. Before Experiments (left). Directed refines edges show how code and skill hypotheses have evolved, while dashed links mark code-skill relations awaiting evidence. After Experiments (right). Solid bidirectional works-with edges identify relations supported by measured positive joint gain.
Figure 4: RoboCasa365 Task Suites. (a) Atomic-Seen contains individual manipulation tasks observed during pretraining. (b) Composite-Seen combines multiple skills into tasks represented in pretraining. (c) Composite-Unseen combines skills into tasks absent from pretraining.
Method
Atomic-Seen
Composite-Seen
Composite-Unseen
Overall
π0
34.6
6.1
1.1
14.8
π0.5
39.6
7.1
1.2
16.9
RLDX-1
60.0
21.3
5.0
30.0
WorldDreamer
66.3
26.7
9.0
35.3
Xiaomi-Robotics-1
80.2
57.1
32.1
57.4
Harness VLA
91.3
55.8
40.1
63.6
Table 1: RoboCasa365 Success Rates. Values are percentages. “–” denotes a suite without a reported result for the corresponding method.
Method
Spat-T
Spat-S
Obj-T
Obj-S
Goal-T
Goal-S
L10-T
L10-S
Overall
OpenVLA
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
π0
0.0
0.0
0.0
2.0
0.0
0.0
0.0
0.0
0.3
π0.5
1.0
20.0
1.0
17.0
2.0
38.0
1.0
8.0
11.0
MolmoAct
0.0
0.0
0.0
6.0
0.0
0.0
6.0
0.0
1.5
NORA
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
X-VLA
0.0
0.0
8.0
2.0
9.0
1.0
10.0
0.0
3.8
Table 2: LIBERO-Pro Success Rates. The table aggregates success rates (%) across SPATIAL, OBJECT, GOAL, and LIBERO-10 under instruction-redirection (T) and position-swap (S) perturbations. “–” denotes a cell not applicable to the corresponding method.
Task
SmolVLA
EmbodiedRSI
Oreo to Red Bowl
70.0%
86.7%
Poker Chip Selection ( ×5=100 )
10.0%
90.0%
Stack Two Cakes on a Can of Luncheon Meat
63.3%
76.7%
Glasses Bridge Grasp
60.0%
56.7%
Towel Folding
26.7%
46.7%
Table 3: Real-Robot Success Rates. Success over 30 trials per task on the real robot.
Figure 6: Code-Skill Co-Evolution Ablation. Bars show RoboCasa365 success rates.
Method
Overall Success @ 50 Ep. ↑
Learning AUC ↑
Random Exploration
27.1
0.309
EmbodiedRSI (Ours)
72.9
0.635
Table 4: Learning Efficiency. EmbodiedRSI is compared with Random Exploration.
Figure 7: Representative LIBERO-Pro Learning Curves. Eight-episode rolling success on SPAT-T over 100 physical episodes. (a) Value-of-Information Experiment Selection reaches sustained 100% at episode 48 (dashed line). (b) Random Selection continues to fluctuate.
Variant
Composite-Unseen ↑
Learning AUC ↑
Retrieval-Based
61.7%
0.556
PPO + Task Reward Only
65.2%
0.589
EmbodiedRSI (Ours)
71.3%
0.635
Table 5: Reward-Grounded Memory Learning Ablation. Composite-Unseen and learning AUC on RoboCasa365.