Publish latest step60 Spatial GRPO checkpoint for continued training

Full model and optimizer state; Spatial 386/500 (77.2%). Use step60 for training beyond step60. Includes resume60-to80 instructions and a compatible reproduction entry point.

README.md

@@ -19,6 +19,24 @@ base_model:
 
 # LIBERO-Spatial 双相机 GRPO 复现
 
+## 最新可续训权重：第 60 步
+
+**最新完整策略断点为 `weights/policy/global_step_60/`。第 60 步之后的训练请从这份断点恢复。**
+它包含完整模型权重、优化器状态及 8 个 DCP 分片；真实 LIBERO-Spatial
+10 个任务 × 50 次评估成功率为 **77.2%（386/500）**。
+`latest_policy.json` 记录最新断点、评估结果和续训入口。40、45、50 步用于历史对比。
+
+下载并执行下文的 `restore`、`check` 后，从 60 步继续训练 20 步，到总步数 80：
+
+```bash
+screen -dmS spatial-grpo60-to80 bash -c \
+  'python3 reproduction/portable.py train --resume 60 --target 80 --output "$PWD/runs/resume60_to80"'
+screen -r spatial-grpo60-to80
+```
+
+`--target 80` 表示训练至总步数 80，即新增 20 步；恢复优化器时使用完整断点目录。
+
+
 本发布包对应 π0.5 + Wan 主视角 + Cosmos-WoVR 腕部视角，在另一台 Linux
 服务器上启动八卡 GRPO。训练使用世界模型生成的观测；成功率是在真实 LIBERO
 仿真器中，以固定初始状态、10 个任务 × 50 次试验统计的 success-once。
@@ -30,7 +48,7 @@ base_model:
 - `runtime/`：Python 3.11/3.10 解释器及两个环境实际使用的依赖层。恢复时保持层的
   顺序，不使用包含重复版本的 `pip freeze` 输出来重建环境。
 - `weights/pi05/global_step_12000/`：匹配的 few-shot SFT 初始化权重和统计量。
-- `weights/policy/global_step_40/`、`global_step_50/`：可供原生 RLinf 续训的
+- `weights/policy/global_step_40/`、`global_step_50/`、`global_step_60/`：可供原生 RLinf 续训的
   完整策略权重及 DCP 训练状态。`global_step_45/` 只有完整策略权重，可用于评估
   或模型初始化，不用于恢复优化器状态。
 - `weights/wan_step1000/`、`weights/cosmos_step2000/`：本轮 GRPO 使用的世界模型
@@ -42,7 +60,7 @@ base_model:
 - `evidence/`：评估结果、模型配对元数据、环境版本清单与实验谱系。
 
 参考检查点的真实 Spatial 成功率：40 步 365/500（73.0%），45 步
-370/500（74.0%），50 步 349/500（69.8%）。40 步由前一轮世界模型训练得到；
+370/500（74.0%），50 步 349/500（69.8%），最新 60 步 386/500（77.2%）。40 步由前一轮世界模型训练得到；
 40→60 的本轮续训使用新 Wan 1000 步和新 Cosmos 2000 步。
 
 ## 目标服务器
@@ -78,7 +96,8 @@ python3 reproduction/portable.py check
 
 ## 启动 GRPO
 
-复现本轮续训，使用第 40 步完整策略及新世界模型，继续到第 60 步：
+历史复现示例：使用第 40 步完整策略及新世界模型，继续到第 60 步。
+第 60 步之后的续训使用上面的最新 60 步断点。
 
 ```bash
 screen -dmS spatial-grpo bash -c \
@@ -110,7 +129,7 @@ python3 reproduction/portable.py train --steps 20 --output "$PWD/runs/new20"
 ## 评估发布的策略
 
 ```bash
-python3 reproduction/portable.py eval --checkpoint 45 --output "$PWD/runs/eval45"
+python3 reproduction/portable.py eval --checkpoint 60 --output "$PWD/runs/eval60"
 ```
 
 该入口只做真实 LIBERO 评估，不执行 GRPO 更新；沿用八卡启动器，因此仍会加载

SHA256SUMS

@@ -1,4 +1,4 @@
-23327ed20129378a8ccf866d6c3dbf54efeec6faa9a7017a231abb964b0b58c8  .gitattributes
+d78e440fa543d86631f243eddb10420ec0731d526d9a486a87ed14597a33e7bc  .gitattributes
 738f95dbf4a56003730d8c7ba113dad975a6636a69a8187fc0bc571dcdfdffbe  LICENSES/COSMOS-MODEL-CARD.md
 3bd882474403177ab22e65598d8ab1c1b299804b5bff0019f30c4a01aa405506  LICENSES/COSMOS-POLICY-MODEL-CARD.md
 c71d239df91726fc519c6eb72d318ec65820627232b2f796219e87dcf35d0ab4  LICENSES/COSMOS-TOKENIZER-LICENSE.txt
@@ -10,7 +10,7 @@ b8817d9f58408022ff30bbf1abd39521e7e291c83b33ff04969c45ac46bd6c65  LICENSES/PI05-
 c7c92fee190b604e2d18e27df517aca77871d0c6829d1a58b71521807d3acdb1  LICENSES/README.md
 4bcf87ecfbbb8e07a01b21415a970c8b53a5283bf6872b657040d3f45c9241f7  LICENSES/WAN-SPATIAL-MODEL-CARD.md
 9491f7a2676a19460f8e9d3c625dabd53841fa8fe3db92bf6cd73780dbf0348e  NOTICE
-acd5ca0db1c9d0b2962dc5bd656f2d008773aaf6c0ee9eb68dbb588518dc05f4  README.md
+a233ff8ba459e8daa3ab72c2c2061c6cb2f9a38f682d9cd0be7840b881b59e0b  README.md
 5b119a98ad7824507ddff3c6c7507ca244261ee416dcfa623876702160c580d3  assets/cosmos/libero_dataset_statistics.json
 8a03499676c6c196127c577144b9fd09bb02c30f9caf058adbcef09bb99ad8f5  assets/cosmos/libero_t5_embeddings.pkl
 38071ab59bd94681c686fa51d75a1968f64e470262043be31f7a094e442fd981  assets/cosmos/tokenizer.pth
@@ -29,12 +29,15 @@ b3fb5c1f5311147b0c53c1881c35e5ef4ec723f06c98bc61e411cafd4871e4b6  evidence/cosmo
 d9c2426cd1209ca943b280c3ed217d410cd571e3085fc3945b1a84befea911a2  evidence/evaluations/global_step_40.json
 0d1100288aebcc2a22fab2470b97a45c8e7d728e38b36e6566e4021898af324f  evidence/evaluations/global_step_45.json
 54128b9e3c050b89b1582cd6fac382543c037096fb41d13dcad34cfb8c392e09  evidence/evaluations/global_step_50.json
+06b80579aa0beae3bf11400d3734e6b5ac9e19bb9fab8c7603ad0123b03a27c2  evidence/evaluations/global_step_60.json
+cc963ddd48281aece1af1a1ccafde791879340728d117f0d692f1136ee289e59  evidence/policy_step60_resume.json
 fdc3869dc035263b05b2c4ff9b2f6826282e3a6cc3f4cd268f43cd9a329d2159  evidence/rlinf_environment.json
 b1d0dca8515f82edbd9ffeb1b33f9023e24f23c68a1a644faf23caefedf7c8d1  evidence/world_model_training_manifest.json
-f6b9b63a7950af4bbd696795b0d9fd3c18f2e0131958d19cf5897ed988fbe41d  reproduction/README.zh-CN.md
+f436503a0dd25df2dc327a73c54dd776945cf3d952c30875942e72032d99492d  latest_policy.json
+1a62781c9104b8dfb0a9d2fd869da6ba0886e6fcf4967dc566090bb4873532c5  reproduction/README.zh-CN.md
 f490f560a262e9402230bb0e7f1f965853dc24c4c72b4d6b3941293c40802e0b  reproduction/eval_train.py
 8d0c9116b5b8624aa8dba4f13b76206b7d326a44aa6a9c2b627076c8266b4582  reproduction/launcher.py
-57d3975f3f67a0bdc18f2c52d91dd1486c92f25965e0accccdf407fc95777cd1  reproduction/portable.py
+b6418fea3feeab7d912f0f84c0a9866cade77c000ec0b574d02fe97295a0230f  reproduction/portable.py
 936193ae5fa7432637b050654633496a62455573abd0224dcfb8e118aceb8f39  runtime/cosmos_packages_0.tar.zst
 bbac3dee6879dc41b12082e07400aa6adac52bbf1c4dc8728e53bf8df20ef709  runtime/cosmos_packages_1.tar.zst
 c4940b5a72db60e1de3eb94534034d4d96be53714acb11044f910c0a8c24f1b2  runtime/python310.tar.zst
@@ -43,6 +46,7 @@ b40be1f29418f4fc9e813c1e801bd200f62c5acfe12e0a1aa38b945d8ea7705c  runtime/python
 a659145d90481d3028e12cc780dead4d8340b64b8e9b692bbec7a3185f359719  runtime/rlinf_packages_0.tar.zst
 19114ac59362c80778ca751a7ae8635717ffb64d13c81ab0451c420c14b29bc9  runtime/rlinf_packages_1.tar.zst
 574001e4ec917eaba32464fa65f81a3257b8691f9b41ea351c19ceff8e552175  source/grpo_spatial_source.tar.zst
+08e1648ba07ce8429bc2faee49fe9a5517fde7d81eae94441e91a07872c91319  source/step60_reproduction_update.tar.zst
 b4a01639bdac0e98073499bbb3a84cd29b2274dba80fa286f0c7ee27bf75b949  weights/cosmos_step2000/checkpoint_manifest.json
 7110ea589cfb2dabf4c90941cf5fc5c5d23fdf47ab78b7a16b56b5014e7a87bf  weights/cosmos_step2000/iter_000002000/model/.metadata
 01f1eb9e46b123df765f614a18835e86393c3bdf7d6147b7be4453bc551b99e4  weights/cosmos_step2000/iter_000002000/model/__0_0.distcp
@@ -81,4 +85,14 @@ b5fcade036cff609f4a594c97b0b6d48a0dee7384d1f426b1daa107498ac03f7  weights/policy
 e988b633f0b8d1b80cb055704801bbdd0efbb26e125abc2aa2d5ea3558d600d2  weights/policy/global_step_50/actor/dcp_checkpoint/__6_0.distcp
 ee625a1335ce36ad687aa57f35d3863e0aae9d4b0bde48240c97540df57ddf7f  weights/policy/global_step_50/actor/dcp_checkpoint/__7_0.distcp
 d38984568e62a9125b0ab9174fd2f0fa5384342db6786b46ecb18b6ab9a7f589  weights/policy/global_step_50/actor/model_state_dict/full_weights.pt
+1ed27d7cab01656a7901e9ccad9bd7bfd2cbd9ebb2e66e30b6b0ce35bd021160  weights/policy/global_step_60/actor/dcp_checkpoint/.metadata
+4305cd70f5e7ba12940141920a86139cd8be7802f487686608a2959bbcd9d317  weights/policy/global_step_60/actor/dcp_checkpoint/__0_0.distcp
+6f2534769e8f33a5ac1e211fa2f02ab0f9162db96f304597302a41cec78d811f  weights/policy/global_step_60/actor/dcp_checkpoint/__1_0.distcp
+4db813d16ed81cc7e985624f832c55cb5077eea32e9101071874289ef8b84a12  weights/policy/global_step_60/actor/dcp_checkpoint/__2_0.distcp
+2ddce9c502a095cac9a981297d4f39eb2853fbd2558040747bf57150381be382  weights/policy/global_step_60/actor/dcp_checkpoint/__3_0.distcp
+69c393b1dc1c6ebadab2acd9da6f2ccc2faf85de0c440db9002deb60f5b75db0  weights/policy/global_step_60/actor/dcp_checkpoint/__4_0.distcp
+f59e42610adf3a01bcf81e7e753054d42d7243661bb1998c427fbb517b0a4597  weights/policy/global_step_60/actor/dcp_checkpoint/__5_0.distcp
+bb83c397885457716c7347f6a978750c57c38c64eecbd535247ed7e6005e1a4f  weights/policy/global_step_60/actor/dcp_checkpoint/__6_0.distcp
+a10bc4924f97525a7eb39c30c9b4437c999524a215006ea49b193deb641ee3b8  weights/policy/global_step_60/actor/dcp_checkpoint/__7_0.distcp
+1160b6e7cb3c135453a68c6491119a061d1ee5b9a7e63afcba25d60be04b7b37  weights/policy/global_step_60/actor/model_state_dict/full_weights.pt
 ed5ae2b151c35b6b509824c829dce196cff1a0ee4d422130a3ad0acf8b45cc47  weights/wan_step1000/model-00001.safetensors

latest_policy.json

@@ -0,0 +1,24 @@
+{
+  "schema": "spatial-grpo-latest-policy-v1",
+  "global_step": 60,
+  "checkpoint": "weights/policy/global_step_60",
+  "full_training_state": true,
+  "optimizer_state": true,
+  "dcp_shards": 8,
+  "parent_policy_step": 55,
+  "world_models": {
+    "wan_step": 1000,
+    "cosmos_step": 2000
+  },
+  "evaluation": {
+    "suite": "libero_spatial",
+    "episodes": 500,
+    "successes": 386,
+    "success_rate": 0.772,
+    "protocol": "10 tasks x 50 trials",
+    "evidence": "evidence/evaluations/global_step_60.json"
+  },
+  "completed_utc": "2026-10-25T13:18:28Z",
+  "recommended_resume_step": 60,
+  "continuation_example": "python3 reproduction/portable.py train --resume 60 --target 80 --output \"$PWD/runs/resume60_to80\""
+}

release_manifest.json

@@ -7,8 +7,8 @@
   "files": [
     {
       "path": ".gitattributes",
-      "size": 637,
-      "sha256": "23327ed20129378a8ccf866d6c3dbf54efeec6faa9a7017a231abb964b0b58c8"
+      "size": 734,
+      "sha256": "d78e440fa543d86631f243eddb10420ec0731d526d9a486a87ed14597a33e7bc"
     },
     {
       "path": "LICENSES/COSMOS-MODEL-CARD.md",
@@ -67,8 +67,8 @@
     },
     {
       "path": "README.md",
-      "size": 6679,
-      "sha256": "acd5ca0db1c9d0b2962dc5bd656f2d008773aaf6c0ee9eb68dbb588518dc05f4"
+      "size": 7625,
+      "sha256": "a233ff8ba459e8daa3ab72c2c2061c6cb2f9a38f682d9cd0be7840b881b59e0b"
     },
     {
       "path": "assets/cosmos/libero_dataset_statistics.json",
@@ -160,6 +160,16 @@
       "size": 39862,
       "sha256": "54128b9e3c050b89b1582cd6fac382543c037096fb41d13dcad34cfb8c392e09"
     },
+    {
+      "path": "evidence/evaluations/global_step_60.json",
+      "size": 39791,
+      "sha256": "06b80579aa0beae3bf11400d3734e6b5ac9e19bb9fab8c7603ad0123b03a27c2"
+    },
+    {
+      "path": "evidence/policy_step60_resume.json",
+      "size": 453,
+      "sha256": "cc963ddd48281aece1af1a1ccafde791879340728d117f0d692f1136ee289e59"
+    },
     {
       "path": "evidence/rlinf_environment.json",
       "size": 22254,
@@ -170,10 +180,15 @@
       "size": 958295,
       "sha256": "b1d0dca8515f82edbd9ffeb1b33f9023e24f23c68a1a644faf23caefedf7c8d1"
     },
+    {
+      "path": "latest_policy.json",
+      "size": 722,
+      "sha256": "f436503a0dd25df2dc327a73c54dd776945cf3d952c30875942e72032d99492d"
+    },
     {
       "path": "reproduction/README.zh-CN.md",
-      "size": 6307,
-      "sha256": "f6b9b63a7950af4bbd696795b0d9fd3c18f2e0131958d19cf5897ed988fbe41d"
+      "size": 7253,
+      "sha256": "1a62781c9104b8dfb0a9d2fd869da6ba0886e6fcf4967dc566090bb4873532c5"
     },
     {
       "path": "reproduction/eval_train.py",
@@ -187,8 +202,8 @@
     },
     {
       "path": "reproduction/portable.py",
-      "size": 14991,
-      "sha256": "57d3975f3f67a0bdc18f2c52d91dd1486c92f25965e0accccdf407fc95777cd1"
+      "size": 15007,
+      "sha256": "b6418fea3feeab7d912f0f84c0a9866cade77c000ec0b574d02fe97295a0230f"
     },
     {
       "path": "runtime/cosmos_packages_0.tar.zst",
@@ -230,6 +245,11 @@
       "size": 290212712,
       "sha256": "574001e4ec917eaba32464fa65f81a3257b8691f9b41ea351c19ceff8e552175"
     },
+    {
+      "path": "source/step60_reproduction_update.tar.zst",
+      "size": 13965,
+      "sha256": "08e1648ba07ce8429bc2faee49fe9a5517fde7d81eae94441e91a07872c91319"
+    },
     {
       "path": "weights/cosmos_step2000/checkpoint_manifest.json",
       "size": 1095,
@@ -420,6 +440,56 @@
       "size": 8526557492,
       "sha256": "d38984568e62a9125b0ab9174fd2f0fa5384342db6786b46ecb18b6ab9a7f589"
     },
+    {
+      "path": "weights/policy/global_step_60/actor/dcp_checkpoint/.metadata",
+      "size": 1570926,
+      "sha256": "1ed27d7cab01656a7901e9ccad9bd7bfd2cbd9ebb2e66e30b6b0ce35bd021160"
+    },
+    {
+      "path": "weights/policy/global_step_60/actor/dcp_checkpoint/__0_0.distcp",
+      "size": 1272885561,
+      "sha256": "4305cd70f5e7ba12940141920a86139cd8be7802f487686608a2959bbcd9d317"
+    },
+    {
+      "path": "weights/policy/global_step_60/actor/dcp_checkpoint/__1_0.distcp",
+      "size": 1268627329,
+      "sha256": "6f2534769e8f33a5ac1e211fa2f02ab0f9162db96f304597302a41cec78d811f"
+    },
+    {
+      "path": "weights/policy/global_step_60/actor/dcp_checkpoint/__2_0.distcp",
+      "size": 1271441317,
+      "sha256": "4db813d16ed81cc7e985624f832c55cb5077eea32e9101071874289ef8b84a12"
+    },
+    {
+      "path": "weights/policy/global_step_60/actor/dcp_checkpoint/__3_0.distcp",
+      "size": 1265552157,
+      "sha256": "2ddce9c502a095cac9a981297d4f39eb2853fbd2558040747bf57150381be382"
+    },
+    {
+      "path": "weights/policy/global_step_60/actor/dcp_checkpoint/__4_0.distcp",
+      "size": 1267590646,
+      "sha256": "69c393b1dc1c6ebadab2acd9da6f2ccc2faf85de0c440db9002deb60f5b75db0"
+    },
+    {
+      "path": "weights/policy/global_step_60/actor/dcp_checkpoint/__5_0.distcp",
+      "size": 1786989589,
+      "sha256": "f59e42610adf3a01bcf81e7e753054d42d7243661bb1998c427fbb517b0a4597"
+    },
+    {
+      "path": "weights/policy/global_step_60/actor/dcp_checkpoint/__6_0.distcp",
+      "size": 2088971714,
+      "sha256": "bb83c397885457716c7347f6a978750c57c38c64eecbd535247ed7e6005e1a4f"
+    },
+    {
+      "path": "weights/policy/global_step_60/actor/dcp_checkpoint/__7_0.distcp",
+      "size": 1559511013,
+      "sha256": "a10bc4924f97525a7eb39c30c9b4437c999524a215006ea49b193deb641ee3b8"
+    },
+    {
+      "path": "weights/policy/global_step_60/actor/model_state_dict/full_weights.pt",
+      "size": 8526557492,
+      "sha256": "1160b6e7cb3c135453a68c6491119a061d1ee5b9a7e63afcba25d60be04b7b37"
+    },
     {
       "path": "weights/wan_step1000/model-00001.safetensors",
       "size": 10050753552,
@@ -476,9 +546,14 @@
       "path": "source/grpo_spatial_source.tar.zst",
       "destination": ".",
       "sha256": "574001e4ec917eaba32464fa65f81a3257b8691f9b41ea351c19ceff8e552175"
+    },
+    {
+      "path": "source/step60_reproduction_update.tar.zst",
+      "destination": ".",
+      "sha256": "08e1648ba07ce8429bc2faee49fe9a5517fde7d81eae94441e91a07872c91319"
     }
   ],
-  "total_bytes": 106009020527,
+  "total_bytes": 126318775207,
   "reference": {
     "suite": "libero_spatial",
     "resume_policy_step": 40,
@@ -500,7 +575,14 @@
       "50": {
         "success": 349,
         "episodes": 500
+      },
+      "60": {
+        "success": 386,
+        "episodes": 500
       }
-    }
-  }
+    },
+    "latest_policy_step": 60,
+    "recommended_resume_step": 60
+  },
+  "updated_utc": "2026-10-26T08:43:18.347958+00:00"
 }

reproduction/README.zh-CN.md

@@ -1,5 +1,23 @@
 # LIBERO-Spatial 双相机 GRPO 复现
 
+## 最新可续训权重：第 60 步
+
+**最新完整策略断点为 `weights/policy/global_step_60/`。第 60 步之后的训练请从这份断点恢复。**
+它包含完整模型权重、优化器状态及 8 个 DCP 分片；真实 LIBERO-Spatial
+10 个任务 × 50 次评估成功率为 **77.2%（386/500）**。
+`latest_policy.json` 记录最新断点、评估结果和续训入口。40、45、50 步用于历史对比。
+
+下载并执行下文的 `restore`、`check` 后，从 60 步继续训练 20 步，到总步数 80：
+
+```bash
+screen -dmS spatial-grpo60-to80 bash -c \
+  'python3 reproduction/portable.py train --resume 60 --target 80 --output "$PWD/runs/resume60_to80"'
+screen -r spatial-grpo60-to80
+```
+
+`--target 80` 表示训练至总步数 80，即新增 20 步；恢复优化器时使用完整断点目录。
+
+
 本发布包对应 π0.5 + Wan 主视角 + Cosmos-WoVR 腕部视角，在另一台 Linux
 服务器上启动八卡 GRPO。训练使用世界模型生成的观测；成功率是在真实 LIBERO
 仿真器中，以固定初始状态、10 个任务 × 50 次试验统计的 success-once。
@@ -11,7 +29,7 @@
 - `runtime/`：Python 3.11/3.10 解释器及两个环境实际使用的依赖层。恢复时保持层的
   顺序，不使用包含重复版本的 `pip freeze` 输出来重建环境。
 - `weights/pi05/global_step_12000/`：匹配的 few-shot SFT 初始化权重和统计量。
-- `weights/policy/global_step_40/`、`global_step_50/`：可供原生 RLinf 续训的
+- `weights/policy/global_step_40/`、`global_step_50/`、`global_step_60/`：可供原生 RLinf 续训的
   完整策略权重及 DCP 训练状态。`global_step_45/` 只有完整策略权重，可用于评估
   或模型初始化，不用于恢复优化器状态。
 - `weights/wan_step1000/`、`weights/cosmos_step2000/`：本轮 GRPO 使用的世界模型
@@ -23,7 +41,7 @@
 - `evidence/`：评估结果、模型配对元数据、环境版本清单与实验谱系。
 
 参考检查点的真实 Spatial 成功率：40 步 365/500（73.0%），45 步
-370/500（74.0%），50 步 349/500（69.8%）。40 步由前一轮世界模型训练得到；
+370/500（74.0%），50 步 349/500（69.8%），最新 60 步 386/500（77.2%）。40 步由前一轮世界模型训练得到；
 40→60 的本轮续训使用新 Wan 1000 步和新 Cosmos 2000 步。
 
 ## 目标服务器
@@ -59,7 +77,8 @@ python3 reproduction/portable.py check
 
 ## 启动 GRPO
 
-复现本轮续训，使用第 40 步完整策略及新世界模型，继续到第 60 步：
+历史复现示例：使用第 40 步完整策略及新世界模型，继续到第 60 步。
+第 60 步之后的续训使用上面的最新 60 步断点。
 
 ```bash
 screen -dmS spatial-grpo bash -c \
@@ -91,7 +110,7 @@ python3 reproduction/portable.py train --steps 20 --output "$PWD/runs/new20"
 ## 评估发布的策略
 
 ```bash
-python3 reproduction/portable.py eval --checkpoint 45 --output "$PWD/runs/eval45"
+python3 reproduction/portable.py eval --checkpoint 60 --output "$PWD/runs/eval60"
 ```
 
 该入口只做真实 LIBERO 评估，不执行 GRPO 更新；沿用八卡启动器，因此仍会加载

reproduction/portable.py

@@ -278,7 +278,7 @@ def run(root, args):
         elif args.resume:
             checkpoint = root / 'weights/policy' / f'global_step_{args.resume}'
             if not (checkpoint / 'actor/dcp_checkpoint/.metadata').is_file():
-                raise ValueError('Resume needs a full DCP checkpoint; use policy step40 or step50')
+                raise ValueError('Resume needs a full DCP checkpoint; use policy step40, step50 or step60')
             if target <= args.resume:
                 raise ValueError('The target must exceed the resumed step')
             command += [f'runner.resume_dir={checkpoint}']
@@ -298,8 +298,8 @@ def main():
     parser.add_argument('command', choices=['restore', 'verify', 'check', 'train', 'eval'])
     parser.add_argument('--root', type=Path, default=Path(__file__).resolve().parents[1])
     parser.add_argument('--quick', action='store_true', help='Only check sizes, without hashing payloads')
-    parser.add_argument('--resume', type=int, choices=[40, 50])
-    parser.add_argument('--checkpoint', type=int, choices=[40, 45, 50], default=45)
+    parser.add_argument('--resume', type=int, choices=[40, 50, 60])
+    parser.add_argument('--checkpoint', type=int, choices=[40, 45, 50, 60], default=60)
     parser.add_argument('--steps', type=int, default=20)
     parser.add_argument('--target', type=int)
     parser.add_argument('--output', type=Path)