bag100/action-atlas-pi05-activations

Pi0.5 Rollouts Directory

Last Updated: 2026-02-05 Model: Pi0.5 (Physical Intelligence) - NOT OpenVLA

VLA interpretability rollout data for causal intervention experiments on Pi0.5.

Note: This is Pi0.5 data. OpenVLA experiments are in /dev/shm/rollouts/ (currently synced to this machine with same structure but different model).

---

Key Findings

1. PaliGemma Activations Dominate Behavior

Finding: PaliGemma layer activations completely override language prompt guidance. The robot executes whatever behavior is encoded in the activations, regardless of the text instruction.

Evidence	Result
Same-scene injection	cos_to_source = 0.999 when injecting source activations
Cross-task injection	0% success - activations transfer behavior, not intent
Cross-prompt baseline	GOAL suite: 0% success with wrong prompt (tasks too different)

2. L0 Is Sufficient for Steering

Finding: Injecting only PaliGemma layer 0 produces nearly identical results to injecting all 18 layers.

Injection Type	Same-Scene Success	Cross-Task Success
pali_L0	32.5%	0.7%
pali_ALL	35.0%	0.7%

This suggests task-specific information is encoded very early in the PaliGemma pathway.

3. Same-Scene Steering Works (+23-26%)

Finding: Within the same scene, injecting correct-task activations improves success by 23-26 percentage points.

Condition	Success Rate
Alt prompt, no injection	9.2%
Alt prompt + inject baseline L0	32.5%
Alt prompt + inject baseline ALL	35.0%
Baseline (correct prompt)	83.3%

4. Cross-Seed Transfer is Robust (91%)

Finding: The model generalizes well across random seeds without intervention.

Condition	Success Rate
Source seeds baseline	93.3%
Target seeds (no inject)	88.9%
Target seeds + inject	87.8-88.9%

Injection provides no benefit - the model already generalizes.

5. Cross-Task Transfer Fails Completely

Finding: Activations are task-specific and do not generalize across different tasks.

Suite	Baseline	With Injection
GOAL	91.9%	0.7-2.2%
SPATIAL	77.3%	4.5-15.9%
LIBERO_10	62.5%	0%

SPATIAL suite shows higher tolerance because tasks share structure (only location differs).

6. Language Is Largely Ignored

Finding: From counterfactual prompting experiments, the model often ignores language.

• 3,396+ episodes across 3 suites show minimal prompt sensitivity
• Null prompts sometimes outperform correct prompts (40% vs 60%)
• Motor programs appear scene-grounded, not language-grounded

7. Early Visual Commitment

Finding: The model commits to a motor program early in the episode.

Perturbation	full	early	late
h_flip	0%	15%	85%
v_flip	10%	15%	85%
rotate_15	50%	50%	85%
grayscale	85%	80%	85%

Late perturbations are tolerated because the model has already committed to its plan.

---

Directory Structure

pi05_rollouts/                    (128 GB total)
├── counterfactual/               # Prompt manipulation (118 GB)
│   ├── object/                   # 1,496 episodes, 5 seeds
│   ├── spatial/                  # 353 episodes + xijia
│   └── goal/                     # 380 episodes + xijia
│
├── cross_task_goal/              # 138 task pairs, 3 seeds (731 MB)
├── cross_task_spatial/           # 23 task pairs (77 MB)
├── cross_task_10/                # 24 task pairs (44 MB)
│
├── vision_perturbation/          # 25+ perturbation types (2 GB)
├── vision_perturbation_xijia/    # Additional runs (350 MB)
│
├── transfer_20260130/            # Cross-scene, saliency, temporal (855 MB)
├── cheng_libero10_5-9_temporal/  # Long-horizon tasks (585 MB)
│
├── pertoken_fast/                # Per-token SAE activations (13 GB)
├── archive/                      # Legacy data + failed runs (6.8 GB)
├── baseline/                     # Baseline control runs
├── logs/                         # Experiment logs
└── DATA_INVENTORY.md             # Detailed inventory

---

Data Summary

Experiment Type	Size	Episodes/Pairs	Key Finding
Counterfactual prompting	118 GB	2,229 episodes	Language ignored
Cross-task transfer	850 MB	360 pairs	0% transfer success
Vision perturbation	2.4 GB	2,000+ experiments	Task-dependent robustness
Temporal perturbation	1 GB	1,500+ experiments	Early commitment
Cross-scene injection	34 MB	68 experiments	0% adaptation

---

Loading Data

import json
import numpy as np
from pathlib import Path

BASE = Path("/data/robotsteering/pi05_rollouts")

# Load counterfactual metadata
with open(BASE / "counterfactual/object/metadata.jsonl") as f:
    episodes = [json.loads(line) for line in f]

# Load layer activations
data = np.load(BASE / "counterfactual/object/expert_layers/layer17/baseline_baseline_0_s42.npz")
activations = data["activations"]  # Shape: [n_steps, hidden_dim]

# Load vision perturbation results
with open(BASE / "transfer_20260130/vision_perturbation/comprehensive/comprehensive_results_20260129_222156.json") as f:
    results = json.load(f)

---

LIBERO Benchmark Suites

Suite	Tasks	Description	Pi0.5 Trained?
libero_object	10	"pick up X and place it in the basket"	✅ Yes
libero_spatial	10	"pick up the black bowl [spatial] and place on plate"	✅ Yes
libero_goal	10	Various kitchen goals (open drawer, turn on stove, etc.)	✅ Yes
libero_10	10	Complex multi-object manipulation	✅ Yes
libero_90	90	Extended benchmark	❌ No

---

Provenance

Data collected from 4 machines (all RTX 4090/5090):

• OMEN (5090): Primary counterfactual/object, temporal perturbations
• XIJIA (4090): Spatial and goal runs (prefixed xijia_*)
• BRYCE (4090): Goal temporal perturbations
• CHENG (4090): LIBERO-10 temporal experiments

Synced to AISM cluster: 2026-01-30

---

See Also

• DATA_INVENTORY.md - Detailed file inventory and counts
• FINDINGS_SUMMARY.md - All validated findings with evidence
• TODO.md - What's still needed for RSS 2026
• plan.md - Methodology from interpretability papers