Robotics
LeRobot
Safetensors
lingbot_va
maximellerbach HF Staff commited on
Commit
9b14f14
·
verified ·
1 Parent(s): 10ab194

Add LingBotActionReferenceStep to the preprocessor

Browse files

Publishes observation.action_zero_norm = norm(0), which the policy uses to re-anchor the action history it feeds back into its KV cache. Under use_relative_actions each chunk is anchored on its own start state, so a verbatim feedback made the action stream reset to zero displacement at every chunk boundary (mean 0.22, p95 0.91 normalized units) and the model double-counted the displacement, ~7 deg/chunk.

Weights are unchanged; this is a preprocessor-only change. Requires a lerobot build that registers `lingbot_action_reference_processor`.

policy_preprocessor.json CHANGED
@@ -86,6 +86,26 @@
86
  },
87
  "state_file": "policy_preprocessor_step_3_normalizer_processor.safetensors"
88
  },
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
89
  {
90
  "registry_name": "device_processor",
91
  "config": {
 
86
  },
87
  "state_file": "policy_preprocessor_step_3_normalizer_processor.safetensors"
88
  },
89
+ {
90
+ "registry_name": "lingbot_action_reference_processor",
91
+ "config": {
92
+ "eps": 1e-08,
93
+ "features": {
94
+ "action": {
95
+ "type": "ACTION",
96
+ "shape": [
97
+ 16
98
+ ]
99
+ }
100
+ },
101
+ "norm_map": {
102
+ "VISUAL": "IDENTITY",
103
+ "STATE": "IDENTITY",
104
+ "ACTION": "QUANTILES"
105
+ }
106
+ },
107
+ "state_file": "policy_preprocessor_step_4_lingbot_action_reference_processor.safetensors"
108
+ },
109
  {
110
  "registry_name": "device_processor",
111
  "config": {
policy_preprocessor_step_4_lingbot_action_reference_processor.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:15778c80a00797e2e1814ded48241146bf9f5e8825a82b678ea4d1f788398259
3
+ size 9400